DEVELOPMENT OF A DRIVER BEHAVIOR BASED ACTIVE COLLISION AVOIDANCE SYSTEM.

Transcription

1 DEVELOPMENT OF A DRIVER BEHAVIOR BASED ACTIVE COLLISION AVOIDANCE SYSTEM. DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree, Doctor of Philosophy in the Graduate School of the Ohio State University By Joshua L. Every, M.S.M.E. Graduate Program in Mechanical Engineering The Ohio State University 2015 Doctoral Examination Committee: Professor Dennis A. Guenther, Advisor Professor Gary J. Heydinger Professor Ahmet Kahraman Professor Junmin Wang

2 Copyright by Joshua L. Every 2015

3 ABSTRACT Modern passenger and commercial vehicles share many of the same safety systems. Advanced Cruise Control, Anti-lock Brakes and Electronic Stability Control have all been shown to be an effective means of improving safety on both classes of vehicles. Dynamic Brake Support (DBS) is a system which has been implemented successfully on passenger cars but no record of implementation on heavy vehicles has been found. This is largely due to the belief that commercial vehicle drivers, as professionals, apply the brakes more effectively than passenger car drivers, and therefore do not need this system. This document presents a multi-point study of the applicability of DBS to commercial vehicles. Beginning with analyzing commercial vehicle driver braking behavior to show that commercial vehicle driver braking behavior is fundamentally similar to passenger car driver behavior. Therefore, systems that assist passenger car drivers should also assist commercial vehicle drivers. Next, a revised method of braking behavior analysis is proposed to better characterize this behavior and model it stochastically. Based on data indicating that this system could be effective, commercial vehicle driver braking behavior was evaluated to show that braking behavior in emergency situations could be reliably distinguished from behavior in non-emergency situations. This is important in that it allows the system to act only in situations in which it is ii

4 necessary. Lastly a prototype DBS system is developed and is shown to be effective at reducing vehicle stopping distance and collision velocity in situations in which the vehicle cannot stop. iii

5 To my wife for her support throughout this process, and to my family for helping me to believe through the bad days and to celebrate the good. iv

6 ACKNOWLEDGMENTS I have been fortunate in my time as a student to work with, learn from, and laugh with a group of fantastic engineers and wonderful people. I would like to begin by thanking my advisor Dr. Guenther, without his insight and encouragement I never would have considered, or been able to pursue, this path. In my time at The Ohio State University I have had the opportunity to learn a from group of great instructors. I would like to thank the members of my committee Dr. Guenther, Dr. Heydinger, Dr. Kahraman and Dr. Wang for their service in advising my work, as well as in the classroom. Throughout my career as a graduate student I have worked with many people to which I owe thanks for their friendship and advice both in research and in life. Amongst these individuals I would like to thank: my fellow graduate students, Scott Zagorski, Sughosh Rao and Sage Wolfe; I would also like to thank Anmol Sidhu, Dale Andreatta, Al Dunn and the team at SEA for their assistance in the research for my Master s Thesis, and throughout graduate school; furthermore, I would like to thank Kamel Salaani, Frank Barickman, Devin Elsasser, Riley Garrott and the team at TRC and VRTC for their contributions to my dissertation research. v

7 VITA Born in Dover, Ohio B.S. in Mechanical Engineering, The Ohio State University, Columbus, Ohio M.S. in Mechanical Engineering, The Ohio State University, Columbus, Ohio 2011-Present Graduate Research Associate, Department of Mechanical and Aerospace Engineering, The Ohio State University, Columbus, Ohio PUBLICATIONS 1. Every, J., Heydinger, G., Guenther, D., Sidhu, A. et al., Design Challenges in the Development of a Large Vehicle Inertial Measurement System, SAE Int. J. Commer. Veh. 7(1):1-7, 2014, doi: / Every, J., Salaani, M., Barickman, F., Elsasser, D. et al., Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh. 7(2): , 2014, doi: / Every, J., Guenther, D., Heydinger, G., A Method of Frequency Content Based vi

8 Analysis of Driving Braking Behavior, SAE Int. J. Commer. Veh. 8(1):2015, doi: / FIELDS OF STUDY Major Field: Mechanical Engineering vii

9 TABLE OF CONTENTS Abstract Dedication Acknowledgments Vita ii iv v vi List of Figures xii List of Tables xvii CHAPTER PAGE 1 Introduction Motivation Outline of Work Presented in this Document Literature Review Anti-Lock Brake Systems Collision Imminent Braking and Forward Collision Warning Systems Brake Assistance and Dynamic Brake Support Systems Driver Behavior Analysis for Development of Advanced Collision Avoidance Systems Future Collision Avoidance Systems Incorporating Braking and Steering Conclusions from Literature Review Chapter 1 References Driver Behavior in Collision Imminent Situations Introduction Background of Analysis Braking Metrics Peak Brake Pressure viii

10 2.3.2 Peak Identification Brake Application Time Steering Behavior Multi-Stage Braking Conclusions Chapter 2 References Frequency Content Based Analysis of Driver Braking Behavior Introduction Motivation Analysis of Panic Events Analysis of Non-Panic Events Methodology Panic Braking Events Non-Panic Braking Events Analysis of a Sample Dataset Panic Scenarios and Results Non-Panic Scenarios and Results Conclusions Chapter 3 References Stochastic Brake Application Profile Generation Introduction Sampling Generating Pseudo-Random Samples from a Given Univariate Distribution Generating Pseudo-Random Samples from Any Univariate Distribution Stochastic Brake Application Time Generation Panic Braking Non-Panic Braking Summary of Results Possible Methods of Brake Application and Brake Pulse Profile Generation Conventional Methods for Profile Generation Frequency Content Based Approach to Stochastic Brake Profile Generation Multivariate Generation of Random Samples Example Stochastic Braking Profiles Generated Using This Method Panic Braking Profiles Non-Panic Braking Profiles Conclusions ix

11 Chapter 4 References Development of a Method for Detecting Driver Braking Behavior Consistent with an Emergency Braking Event Introduction Conventional Panic Detection Methodology Conventional Methods of Detecting Driver Intention by Determining When Braking Behavior is Consistent with a Panic Response Conventional Methods of Detecting Driver Intention by Determining When Braking Behavior is Consistent with a Non-Panic Response Panic Detection Methodology Utilizing Multiple Metrics Simultaneously Detection of Driver Intention by Combining Non-Panic Behavior Data for Multiple Metrics Conclusions Continuing Work Chapter 5 References Development and Simulation of a Prototype DBS Algorithm for Heavy Vehicles Introduction Overview of DBS Algorithm used for Simulation Testing Selection of TTC Threshold Panic Braking Behavior Detection Methodology Simulation Diagrams and Code in the MATLAB/ Simulink Environment Vehicle Model Used for Evaluation of the Prototype DBS System Simulation Results and Comparison between Different Combinations of Systems and Thresholds Evaluation of the Influence of TTC on System Effectiveness Comparison of Simulation Results using DBS Systems with Varying Detection Methods DBS System Simulation, Conclusions and Continuing Work Proposal of an Expanded Forward Collision Avoidance and Mitigation System with Features of FCW, CIB and DBS. 132 Chapter 6 References Contributions to the Engineering Community and Continuing Work Contributions Continuing Work x

12 7.2.1 Evaluation of DBS using HIL Simulation to Evaluate its Effectiveness in Conjunction with Other Safety Systems Use of Stochastic Brake Pedal Force Generation in Evaluation of Various DBS Activation Methods and Thresholds Bibliography APPENDICES A Explanation of Scenarios Presented to Drivers and Vehicle Configurations for NADS Simulations A.1 Scenario Design Appendix A References B Kruskal-Wallis Test Box Plots for Brake Application Time C Simulink Block Diagrams and MATLAB code used for DBS testing C.1 Pedal Force Profile Generation Subsystem C.1.1 Fundamental Slope MATLAB Function Code C.1.2 Profile Deviation Generation MATLAB Function Code C.2 DBS System Subsystem C.2.1 DBS System Activation Subsystem C.2.2 DBS Enable MATLAB Function Code C.2.3 DBS Pressure Control Subsystem C.2.4 TreadleValve MATLAB Function Code C.2.5 DBS System Pressure MATLAB Function Code C.3 MATLAB Code Used to Setup and Run DBS Simulations xi

13 LIST OF FIGURES FIGURE PAGE 2.1 Brake force to treadle pressure map for NADS simulation Example plot of method of peak identification (Data from [1]) Histogram of brake application time for all brake systems and different scenarios (0.15 second bin width) (Data from [1]) Example plot of method of pause identification (Data from [1]) Plots of various brake force profiles that exhibit multi-stage behavior (Data from [1]) Example plot of a brake force application profile parameterized using a two-stage approach Plot of an example brake application profile plotted with the fundamental slope (Data from [3]) Plot of the profile deviation from the fundamental slope for an example brake application profile (Data from [3]) Example plot of curve fitting the Fourier series to the data while varying the number of Fourier coefficients used (Data from [3]) Plot of an example non-panic braking profile (Data from [3]) Plot of an example non-panic braking profile with Fourier series fit (Data from [3]) Plot of mean and 95% limits of the Fourier coefficients from panic braking v. frequency in terms of multiples of the fundamental (Data from [3]) xii

14 3.8 Plot of mean and 95% limits of the Fourier coefficients from non-panic braking v. frequency in terms of multiples of the fundamental (Data from [3]) Example plot of sampling an exponential distribution using the CDF Plot of empirical CDF of brake application time for panic braking (Data from [1]) Plot of empirical CDF of brake pulse duration for non-panic braking (Data from [1]) Example plot of parameterized multi-staging profile Plot of various braking profiles which demonstrate multi-staging, used to show variability (Data from [1]) Example plot of parameterized non-panic brake profiles Plot of parameterized fundamental slope Flow chart of the process of mapping data from the native distribution to the standard normal distribution. Gray boxes represent data mapping operations. Red blocks represent datasets Flow chart of the process of mapping pseudo-random samples from the multivariate normal distribution to the native distributions. Gray boxes represent data mapping operations. Red blocks represent datasets Example plot of brake application profiles with stochastic frequency content for 8 participants using 10 frequency coefficients Example plot of non-panic brake pulse profiles with stochastic frequency content for 5 participants using 100 frequency coefficients Plot of the empirical CDF of maximum brake pedal force from drivers in panic situations (Data from [1]) Plot of the empirical CDF of maximum brake pedal force rate from drivers in panic situations (Data from [1]) Plot of the empirical CDF of maximum brake pedal force from drivers in non-panic situations (Data from [1]) Plot of the empirical CDF of maximum brake pedal force rate from drivers in non-panic situations (Data from [1]) xiii

15 5.5 Plot of the empirical CDFs of maximum brake pedal force from drivers in both panic and non-panic situations (Data from [1]) Plot of the empirical CDFs of maximum brake pedal force rate from drivers in both panic and non-panic situations (Data from [1]) Plot of the empirical CDFs of maximum PBI from drivers in both panic and non-panic situations (Data from [1]) Block diagram of proposed DBS system architecture Diagram of right incursion scenario Diagram of stopped vehicle scenario Plot of the empirical CDF of TTC at initialization of braking (Data from [1]) Block diagram of panic detection method Top level of the Simulink diagram used for DBS simulation Diagram of the contents of the Pedal Force Profile Generation Block Diagram of the contents of the DBS system block Comparison of brake system pressure with and without DBS (Driver brake profile from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = L AS (256 s) (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = L BS (254 s) (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 8 s (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 6 s (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 4 s (Driver brake profiles from [1]) xiv

16 6.15 Example plot of brake system command pressure, velocity and deceleration, with and without DBS, for a single application profile with a PBI threshold of 99.5 and a TTC threshold of 6 seconds (Driver brake profile from [1]) Plot of the CDFs of stopping distance for the activation methods and threshold sets previously discussed (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 95 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99.5 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99.8 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force rate based activation with a non-panic percentile threshold of 84 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force rate based activation with a non-panic percentile threshold of 94 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force based activation with a non-panic percentile threshold of 90 (Driver brake profiles from [1]) Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force based activation with a non-panic percentile threshold of 96 (Driver brake profiles from [1]) Photograph of NHTSA s heavy vehicle pneumatic HIL system Photograph of the brake pedal force controller attached to the pneumatic HIL system A.1 Road geometry xv

17 A.2 Left incursion scenario diagram A.3 Right incursion scenario diagram A.4 Stopping vehicle scenario diagram A.5 Stopped vehicle scenario diagram B.1 Left incursion, brake application time, Kruskal-Wallis box plot B.2 Right incursion, brake application time, Kruskal-Wallis box plot B.3 Stopping vehicle, brake application time, Kruskal-Wallis box plot B.4 Stopped vehicle, brake application time, Kruskal-Wallis box plot B.5 Scenario comparison, brake application time, Kruskal-Wallis box plot. 160 C.1 Top level of Simulink diagram used for DBS simulation C.2 Diagram of the contents of the Pedal Force Profile Generation Block. 163 C.3 Diagram of the contents of the Frequency Content Based Brake Force Profile Generation Block C.4 Diagram of the contents of the DBS system Block C.5 Diagram of the content of the DBS System Activation Block C.6 Diagram of the contents of the DBS Pressure Control Block xvi

18 LIST OF TABLES TABLE PAGE 2.1 Summary of number of collisions with incursion vehicles (Data from [1]) Mean collision speed and (mean velocity difference) for collisions with simulated vehicles in MPH (Data from [1]) Population distribution of datasets analyzed (Data from [1]) Mean and standard deviation of peak brake pressure for each combination of brake system and scenario (Data from [1]) Percentage of individuals who achieve brake system saturation for each combination of brake system and scenario (Data from [1]) Mean, standard deviation and p-value of brake application time for each combination of brake system and maneuver (Data from [1]) P-values of brake application time comparing each set of brake system type (Data from [1]) Mean and standard deviation of magnitude of mean steer angle for each combination of brake system and maneuver (Data from [1]) Percentage of tests that exhibit multi-stage braking for each combination of scenario and brake system with varying brake pause threshold and minimum pause time (Data from [1]) Summary of number of collisions with incursion vehicles and (number of total tests) - W: with multi-staging; W/O: without multi-staging. Multi-stage braking evaluated with 25% brake pause threshold and 0.5 sec minimum pause time (Data from [1]) xvii

19 2.11 Mean collision speed and (mean velocity difference) for collisions with simulated vehicles in MPH - W: with multi-staging; W/O without multi-staging. Multi-stage braking evaluated with 25% brake pause threshold and 0.5 sec minimum pause time (Data from [1]) Fourier coefficients A 1 through A 10 for the profile deviation fit shown in Figure 3.4 (Data from [3]) Fourier coefficients A 1 through A 10 for the profile fit shown in Figure 3.6 (Data from [3]) Table of PBI panic detection percentage and false detection percentage for varying levels of PBI threshold Table of mean and standard deviation of TTC at initialization of braking, in seconds, and data for the number of profiles with a valid TTC at initialization of braking(data from [1]) Table of threshold sets used to evaluate the influence of the TTC Threshold on system performance Variation in system activation percentage with varying TTC Threshold (Driver brake profiles from [1]) Brake application time results from analysis of variation in TTC threshold (Driver brake profiles from [1]) Stopping distance results from analysis of variation in TTC threshold (Driver brake profiles from [1]) Change in velocity, at various distances after the initialization of braking, results from analysis of variation in TTC threshold (Driver brake profiles from [1]) Table of threshold sets used to evaluate the influence of the panic detection method on system performance Variation in system activation percentage with varying detection method and parameters (Driver brake profiles from [1]) Brake application time results from analysis of variation in detection method and parameters (Driver brake profiles from [1]) Stopping distance results from analysis of variation in detection method and parameters (Driver brake profiles from [1]) xviii

20 6.11 Change in velocity, at various distances after the initialization of braking, results from analysis of variation in detection method and parameters (Driver brake profiles from [1]) xix

21 CHAPTER 1 INTRODUCTION 1.1 Motivation Compact electronic systems have become abundant in all aspects of life, and the design of modern passenger, and commercial, vehicles has been no exception. Applications from wireless tire pressure monitors to electronic stability control have been used to make vehicles safer and more efficient. One area which has seen extensive and broad development is the application of this technology to active collision avoidance and mitigation systems. The majority of systems being developed or used in the field are focused on incidents in which the initial point of contact on the striking vehicle is in the front of the vehicle. This trend of development is supported by accident statistics. In reviewing heavy truck collision statistics provided by the U.S. Department of Transportation [1], of all two-vehicle crashes involving large trucks, the initial impact point of contact on the heavy truck was the front of the vehicle in 62% of cases and the initial point of contact of the other vehicle was the front of the vehicle in 65% of cases. This does not imply that approximately 60% of collisions involving heavy vehicles are head on but it does imply that a large percentage of collisions involve either the front end of a heavy vehicle striking another vehicle or the heavy vehicle being struck by the front end of another vehicle. This indicates that the 1

22 maximum plausible effectiveness of a single forward collision system is activation in approximately 60% of collisions but if systems were implemented globally this percentage could be increased significantly. Furthermore, the technology of collision avoidance is about milliseconds not minutes. Prior research has shown that if driver reaction time could be improved by a half second it is possible that 60% of front end crashes and one third of head on collisions could be avoided [2]. To this end systems that generate seemingly marginal improvements in driver response could result in significant improvements in safety. The development of new and novel safety systems should therefore consider improving driver response as a key factor in improving safety, having the same level of importance as improving vehicle performance or control. 1.2 Outline of Work Presented in this Document Based on this information and the literature reviewed, an investigation of the implementation of Dynamic Brake Support (DBS) on commercial vehicles was conducted. DBS is a system designed to shorten the time it takes a driver to apply the brakes in emergency situations. The presentation of information in this document follows the progression of work in the course of this project. The initial question asked in this work was: do commercial vehicle drivers exhibit multi-stage braking behavior in significant enough numbers for DBS to be effective in assisting these drivers? This question will be answered by the results of the study presented in Chapter 2. Once this behavior had been quantified it was realized that a better method of analyzing braking behavior could be developed. The goal of this method is to extract more data about driver behavior than what is currently possible. This method is presented in Chapter 3. The results of this work led to the question of whether this behavior could be modeled 2

23 stochastically in order to aid in the development and evaluation of any candidate algorithm. The results of developing such a method are presented in Chapter 4. After this behavior had been analyzed and stochastically modeled the natural next step is to develop a method of implementing DBS in these cases. A method of detecting panic braking was developed and the results of this development are presented in Chapter 5. After this algorithm was developed it was tested using TruckSim in order to determine the possible improvement in outcomes when drivers are placed in collision imminent situations. The results of these simulations are presented in Chapter 6. Finally, the contributions of this work to the engineering community are presented in Chapter Literature Review In order to evaluate the improvement in safety possible through the implementation of more advanced safety systems an understanding of the current state of the art is necessary. The following sections provide information on safety systems that have already been implemented, or are being researched for possible use on passenger or commercial vehicles Anti-Lock Brake Systems One of the earliest widespread applications of such a system was Anti-Lock Brake Systems (ABS). Though not explicitly used for collision mitigation the benefits of this system of improved steerability and reduced stopping distance allows drivers to better respond to situations with a high probability of the occurrence of a collision. Studies have confirmed the effectiveness of ABS in assisting passenger car drivers by maintaining the ability to steer while heavily braking, especially on wet roads [3] [4]. ABS has also been shown to have great potential to improve safety for commercial 3

24 vehicles [5]. This potential is supported by field testing of straight-trucks [6] and for articulated vehicles with various system architectures [7]. Though ABS is highly effective at assisting a driver in maintaining vehicle control in an emergency, the fact remains that the effectiveness of such a system is limited by the extent to which the driver applies the brakes. It is well documented that in many cases, drivers of passenger vehicles do not apply sufficient braking in situations in which a collision is possible or imminent. In [3] it was found that in a study of passenger car drivers in emergency situations only 31% of drivers on dry pavement braked sufficiently to either, activate ABS amongst drivers with ABS, or achieved wheel lockup amongst drivers without ABS. Also, in [8] it was found that in the examination of fifty case studies of rear end collisions on single lane roads, indications of vehicle skid were only found in 22% of cases. This indicates that in a majority of cases, 78%, drivers do not input large enough braking forces to saturate the tires. Therefore, drivers do not generally achieve peak vehicle deceleration in these cases. Also, presented in this document is a study of driver behavior in which the authors conclude that, of 82 participants 47% did not input sufficient pedal force when presented with an object jumping-out in front of the test vehicle. The authors characterized this behavior as drivers in both groups (those who do and do not input sufficient pedal force) initially applying the same brake effort. Then some drivers settled into a lower level of constant or decreasing brake force, while others increased to a significantly higher pedal force. This observed behavior agrees with the belief that in an emergency situation drivers often exhibit what is called a two-stage braking process. In [9] Prynne and Martin state that this behavior pattern is a result of humans not having instinctive reactions to situations of vehicle emergency. They explain that this two-stage process is related to the perception of, and reaction to the situation. The first stage is an initial 4

25 reaction to the given situation, in which the driver applies braking at approximately one third of the vehicles braking capacity, which is the limit of normal or nonemergency braking. This option is chosen because it presents a reasonable course of action without eliminating other possible options, such as steering away from the situation, and allows the driver further time to process. The second stage occurs after the driver has had time to process and decide on a course of action. Assuming that braking is chosen, they then increase braking force to a significantly higher value Collision Imminent Braking and Forward Collision Warning Systems State of the art forward collision mitigation technologies which have been introduced recently and are beginning to see widespread commercial availability, are known as Forward Collision Warning (FCW) and Collision Imminent Braking (CIB) systems. The main function of FCW is the issuance of audible or haptic warnings to a driver in situations in which a forward collision is likely to occur in the absence of driver intervention. Whereas the function of CIB is intervention through autonomous braking in cases in which a collision is unavoidable or highly likely. Though CIB can improve safety in some cases it may not activate in situations when a driver is actively controlling the vehicle, even if driver intervention is not sufficient to avoid a collision. Due to these systems having largely independent functionality they are often combined in a system referred to as Forward Collision warning And Mitigation (F-CAM) systems. Most often CIB systems activate based on knowledge of vehicle following distance and determination of when a certain following distance function is violated. This following distance information can be obtained using a wide array of vehicle mounted sensors as discussed in [2]. An example of this type of implementation on commercial 5

26 vehicles is given in [10]. More advanced methods of CIB activation have also been researched as shown in [11]. Once a method of activation has been established these systems can be evaluated to determine the expected improvement in safety though their use. In [12] and [13] the possible improvements in safety via a decreased number of fatalities, and the reduction in the level of injury possible through the use of an F-CAM system, on passenger vehicles, is demonstrated by comparing results from vehicle crash testing. As well, in [14] the results from a study indicating that a collision mitigation braking system could be effective in assisting drivers in as many as 24% of all collisions of passenger cars is presented. Depending on the peak deceleration which can be induced by such a system various levels of performance can be achieved. If a 3 m/s 2 deceleration were possible 45% of frontal impacts could be avoided entirely and 48% could be mitigated. While, if the deceleration possible were increased to 6 m/s 2, 57% of accidents could be avoided, and with an allowable deceleration of 8 m/s 2, 68% of accidents could be avoided. The effectiveness of CIB, FCW and F-CAM systems has also been evaluated for commercial vehicles. In [15] the method of implementation of various safety systems including FCW and CIB on commercial vehicles in Japan is presented. In [16] and [17] the results from a study of the possible effectiveness of FCW, CIB and F- CAM systems, available at the time (2013) on commercial vehicles and the possible improvements in safety that could be realized through the use of more advanced systems, are presented. In these studies the current (2013) generation system is defined as a system incapable of detecting vehicles that are stopped when they enter into the detection range and the system is capable of generating a deceleration of less than 0.4 g. Also, the second and third generation systems are modified when compared to the original system such that, the second generation systems have the 6

27 capability of detecting vehicles that are stopped when they come into the range of the vehicle radar and are capable of generating a deceleration of 0.35 g. The third generation systems have the same abilities as the second generation with the added authority of being able to generate a deceleration of 0.6 g. The results presented indicate that through global usage of a current generation F-CAM system, fatal collision injury severity could be reduced by 24% and injury collision severity could be reduced by 25% for tractor trailers. Also in the scenarios simulated for single unit trucks, fatal collision injury severity could be reduced by 22% and injury collision severity could be reduced by 21% [17]. Through the use of more advanced systems, the reduction in injury severity for tractor trailers in fatal accidents is 44% and for injury accidents is 47% for vehicles using the second generation system. For the third generation system the reduction in injury severity in fatal accidents is 57% and in injury accidents is 54% [16] Brake Assistance and Dynamic Brake Support Systems In order to address the pattern of behavior in which drivers do not input sufficient brake force, in situations in which a collision is probable or imminent, brake assistance systems have been developed and implemented on passenger cars. These systems have been proposed based on both mechanical [18] and electronic [8] activation methods. These systems have been shown to be effective at increasing the average deceleration of drivers in collision imminent situations [8]. Such systems are typically activated based on thresholds of brake pedal force, brake pedal speed or combinations of the two. The main draw back of these systems is that they do not use external sensing to determine if activation is necessary. This means that if a driver issues a panic like brake input when driving with no other vehicles or objects in its path the system may 7

28 activate. This behavior leads to a reduced level of driver acceptance and an increased possibility of other dangers associated with false activation. By combining algorithms developed for brake assistance systems and sensing used in forward collision warning and collision imminent braking systems a system referred to as Dynamic Brake Support (DBS) was created. DBS systems differ from CIB systems in that brake activation does not occur autonomously, but if an input consistent with panic braking is detected, and a vehicle with which a collision is possible is sensed, the DBS system will assist the driver by applying significantly increased brake pressure. This system can activate earlier than CIB and is not limited such that it may not activate when the driver is actively controlling the vehicle. Brake assist is applied based on various methods of detecting when driver braking behavior is consistent with their response to a panic situation. Evaluation of passenger car driver behavior and threshold selection for either pedal speed or pedal force based activations have also been analyzed [19]. The results of this analysis showed that with proper threshold selection either parameter can be effective, with pedal speed having the possibility of activating earlier, but also being more prone to false activations. Also, an advanced method of activation based on monitoring driver foot speed during the accelerator to brake pedal transition has been proposed, but no instance of this being implemented on a publicly available vehicle has been found [20]. Data for effectiveness of both brake assistance and DBS systems, along with an explanation of one manufacturer s method of implementation on passenger vehicles is explained in [21]. Brake assistance systems have also been researched for commercial vehicles. In [22] and [23] a method of applying brake assistance to commercial vehicles is implemented as part of a larger full vehicle safety system. This method uses brake pedal rate in terms of stroke percentage per second to determine if application is consistent with 8

29 panic braking. This system uses a rate threshold with system activation occurring once it is exceeded. The system then remains enabled until the driver has released the brake pedal Driver Behavior Analysis for Development of Advanced Collision Avoidance Systems As systems are being developed which respond to not only external inputs but driver behavior to determine if intervention is necessary, the understanding of driver behavior in collision imminent situations is of increasing importance. In [24] analysis of passenger car driver behavior in collisions is presented. It is stated that in 50% of collisions driver braking does not generate a deceleration greater than 2 m/s 2. Also the level of braking that one would expect to see prior to a collision, greater than 8 m/s 2, is only seen in 1% of cases. This combined with other data presented leads to the conclusion that human error is a factor in 93% of all cases presented. In [25] the results from a study evaluating simulator data from 384 drivers placed in collision imminent situations using the National Advanced Driving Simulator (NADS) are presented. This study found that commercial vehicle drivers in collision imminent situations tend to exhibit multi-stage braking behavior as previously documented for drivers of passenger cars. These results indicate that if a DBS system, that could detect this behavior, were implemented on heavy trucks safety would be improved. In [26] data from a NHTSA Study ([3]) used to evaluate the effectiveness of ABS is reevaluated in order to better understand passenger car driver braking and steering behavior in situations in which a sudden emergency arises. The results of this study found that a majority of subjects, 61%, used braking and steering in their attempt to avoid this particular event. Another significant portion of drivers, 28%, of drivers used 9

30 only braking in these situations. Overall this implies that 89% of drivers employed some braking in their response to a sudden emergency situation. In [27] passenger car drivers braking response to a situation in which they are following a vehicle which begins decelerating is presented. In this study driver behavior was analyzed in order to determine how driver braking behavior varied based on driver perception of how hard braking was expected to be applied. This study presents a method for classifying a scenarios severity based on knowledge of the current range and range rate. Multiple metrics can be used to evaluate the difference in safety in various situations. Two metrics generally used are the Time To Collision (TTC) and headway distance between two vehicles. TTC is defined as Headway divided by relative velocity, which is equivalent to the time until a collision occurs if neither driver intervenes. Comparison of using TTC and headway to evaluate relative risk at various locations in an intersection in presented in [28]. Where it is concluded that, though headway is a more reasonable means of enforcing or instructing safe following distance, TTC is more closely related to the probability of a collision occurring. Furthermore, data for TTC at the onset of braking is estimated for various real world crash scenarios in which the lead vehicle was stopped is presented in [29]. The average TTC at onset of braking for passenger car drivers presented in this study ranged between 1.1 and 1.4 seconds. In [30] Chen and Dai present results from the development of an advanced cruise control system which varies system activation based on perception or driver fatigue. By making the system more active when driver fatigue is sensed the effectiveness of the system can be improved in the cases when it is most necessary. This type of implementation limits the system in cases when the driver is less fatigued, this effect improves driver acceptance. 10

31 1.3.5 Future Collision Avoidance Systems Incorporating Braking and Steering In the evaluation of driver behavior in collision imminent scenarios it has been observed in multiple studies that steering away from a collision is a largely underused tactic of collision avoidance, even though the distance required to make a controlled lane change is generally shorter than that required to stop. In [9] Prynne and Martin observed that a majority of drivers in collision imminent situations react by braking when in many cases only a moderate steering input would be required to avoid the collision. Also observed in this study is that, drivers who steer to avoid a collision have a lower incidence of hitting an obstacle than drivers who brake. As well, drivers who employ both braking and steering have the highest probability of avoiding a collision. In [31] Lechner and Malaterre observed that of the 45% of accidents, in a case study, that could have been avoided by proper driver intervention, the required intervention in 50% of cases was a steering input, while in 80% of cases the drivers solely applied the brakes. The reluctance of drivers to steer or to provide sufficient steering inputs has led to the desire to develop active collision avoidance systems that incorporate steering into the list of possible collision avoidance maneuvers. These systems generally act by autonomously steering the vehicle or by supplementing driver steering inputs. Steering inputs can be supplemented either through the use of supplementing driver steering torque, steer angle, or by implementing some other form of directional control i.e. differential braking. For systems designed to supplement driver input, it is first required to discern the steering intention of a driver, this is generally done based on measured steering inputs. In [32] a method of determining when a driver has used steering inputs which are consistent with an intention to steer the vehicle in a given direction is proposed 11

32 and demonstrated. These parameters are specific to the vehicle and are based on oncenter handling performance and include steering wheel torque, steering wheel angle, lateral acceleration, steering wheel angular speed and vehicle speed. In [33] Choi and Yu propose the use of a steering assist system that aids the driver by applying additional steer torque and differential braking. They state that this method may be more readily accepted by drivers as there is less indication of system activation than in systems which control steering wheel angle. The proposed method of driver assist uses what is referred to as a lane change index to determine the current phase of operation that the driver is pursuing and provides assistance specific to that particular phase. This method of control uses the given available inputs to direct the vehicle toward a path designed to maximize the minimum distance between the vehicles in a given scenario. Results from simulation indicate that this system may be effective in assisting drivers in avoiding frontal collisions. In [34] Yang presents a method of determining the maximum speed for which a vehicle can pass the double lane change maneuver. Though not particularly used for collision avoidance, path optimization is used to determine the path through which the highest speed without instability can be achieved. In [35] the development of a trajectory planning method that can be used with autonomous vehicles for collision avoidance is discussed. The method used in this study is capable of generating a steering path based on knowledge of surrounding traffic and selecting a path that averts a collision. This method is simulated and demonstrated in scenarios of overtaking a vehicle and avoiding a vehicle at an intersection with a perpendicular path of travel. No references have currently been found describing an active steering assistance system being developed for commercial vehicles. 12

33 1.3.6 Conclusions from Literature Review Generally in the development of advanced safety systems, systems are initially implemented on passenger cars and are later applied to commercial vehicles. Also, DBS has been applied to passenger cars and has shown to be effective at improving safety. Based on this information it has been decided that, the implementation of DBS on commercial vehicles should be investigated to determine if the trend of technology migrating from passengers cars to commercial vehicles is also valid in this case. 13

34 Chapter 1 References [1] NHTSAs National Center for Statistics and Analysis, Traffic Safety Facts, 2012 Data, Large Trucks. DOT HS , May [2] P. M. Knoll, B.-J. Schaefer, H. Guettler, M. Bunse, and R. Kallenbach, Predictive Safety Systems - Steps Towards Collision Mitigation, in SAE Technical Paper, no , SAE International, [3] E. N. Mazzae, F. Barickman, G. H. S. Baldwin, and G. Forkenbrock, Driver Crash Avoidance Behavior with ABS in an Intersection Incursion Scenario on Dry Versus Wet Pavement, in SAE Technical Paper, no , SAE International, [4] L. Evans, ABS and Relative Crash Risk Under Different Roadway, Weather, and Other Conditions, in SAE Technical Paper, no , SAE International, [5] R. Emig, H. Goebels, and H. J. Schramm, Antilock Braking Systems (ABS) for Commercial Vehicles - Status 1990 and Future Prospects, in SAE Technical Paper, no , SAE International, [6] S. B. Zagorski and R. L. Hoover, Comparison of ABS Configurations and Their Effects on Stopping Performance and Stability for a Class 8 Straight-Truck, in SAE Technical Paper, no , SAE International, [7] M. L. Shurtz, G. J. Heydinger, D. A. Guenther, and S. B. Zagorski, Effects of ABS Controller Parameters on Heavy Truck Model Braking Performance, in SAE Technical Paper, no , SAE International, [8] H. Yoshida, T. Sugitani, M. Ohta, J. Kizaki, A. Yamamotoz, and K. Shirai, Development of the Brake Assist System, in SAE Technical Paper, no , SAE International, [9] K. Prynne and P. Martin, Braking Behaviour in Emergencies, in SAE Technical Paper, no , SAE International, [10] S. Chakraborty, T. A. Gee, and D. Smedley, Advanced Collision Avoidance Demonstration for Heavy-Duty Vehicles, in SAE Technical Paper, no , SAE International, [11] J. Jansson, J. Johansson, and F. Gustafsson, Decision Making for Collision Avoidance Systems, in SAE Technical Paper, no , SAE International, [12] P. Ruecker, Crash Tests with Automatic Pre-Crash Braking Cars, in SAE Technical Paper, no , SAE International,

35 [13] M. Egelhaaf and P. Rücker, Incidence of Frontal Impact Accidents and Crash Testing of Cars Equipped with Collision Imminent Braking Systems, in SAE Technical Paper, no , SAE International, [14] M. Lindman and E. Tivesten, A Method for Estimating the Benefit of Autonomous Braking Systems Using Traffic Accident Data, in SAE Technical Paper, no , SAE International, [15] H. Enomoto, K. Akiyama, and H. Okuyama, Advanced Safety Technologies for Large Trucks, in SAE Technical Paper, no , SAE International, [16] J. Woodrooffe, D. Blower, C. A. C. Flannagan, S. E. Bogard, P. A. Green, and S. Bao, Automated Control and Brake Strategies for Future Crash Avoidance Systems - Potential Benefits, in SAE Technical Paper, no , SAE International, [17] J. Woodrooffe, D. Blower, C. A. C. Flannagan, S. E. Bogard, and S. Bao, Effectiveness of a Current Commercial Vehicle Forward Collision Avoidance and Mitigation Systems, in SAE Technical Paper, no , SAE International, [18] H. J. Feigel and J. Schonlau, Mechanical Brake Assist - A Potential New Standard Safety Feature, in SAE Technical Paper, no , SAE International, [19] T. Hirose, T. Taniguchi, T. Hatano, K. Takahashi, and N. Tanaka, A Study on the Effect of Brake Assist Systems (BAS), SAE Int. J. Passeng. Cars Mech. Syst., vol. 1, no , pp , [20] S. Kitazawa and Y. Matsuura, Feasibility Study of a Braking Assistant System for Driver Pedal Operation in Emergency Situations, in SAE Technical Paper, no , SAE International, [21] J. J. Breuer, A. Faulhaber, P. Frank, and S. Gleissner, Real world safety benefits of brake assistance systems, in 20th International Technical Conference on the Enhanced Safety of Vehicles (ESV), [22] D. Zhang, C. Zong, S. Yang, and W. Zhao, Development and Verification of Electronic Braking System ECU Software for Commercial Vehicle, in SAE Technical Paper, no , SAE International, [23] J. Han, Z. Changfu, and Z. Weiqiang, Development of a Control Strategy and HIL Validation of Electronic Braking System for Commercial Vehicle, in SAE Technical Paper, no , SAE International,

36 [24] P. M. Knoll, Predictive Safety Systems: Convenience - Collision Mitigation - Collision Avoidance, in SAE Technical Paper, no , SAE International, [25] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [26] G. F. Fowler, R. E. Larson, and L. A. Wojcik, Driver Crash Avoidance Behavior: Analysis of Experimental Data Collected in NHTSAs Vehicle Antilock Brake System (ABS) Research Program, in SAE Technical Paper, no , SAE International, [27] W. G. Najm and D. L. Smith, Modeling Driver Response to Lead Vehicle Decelerating, in SAE Technical Paper, no , SAE International, [28] K. Vogel, A comparison of headway and time to collision as safety indicators, Accident analysis & prevention, vol. 35, no. 3, pp , [29] K. D. Kusano and H. Gabler, Method for Estimating Time to Collision at Braking in Real-World, Lead Vehicle Stopped Rear-End Crashes for Use in Pre- Crash System Design, SAE Int. J. Passeng. Cars Mech. Syst., vol. 4, no , pp , [30] L.-k. Chen, C.-c. Dai, and M.-f. Luo, Investigation of a driver-oriented adaptive cruise control system, International Journal of Vehicle Design, vol. 66, no. 1, pp , [31] D. Lechner and G. Malaterre, Emergency Manuever Experimentation Using a Driving Simulator, in SAE Technical Paper, no , SAE International, [32] T. Hong, J. Kwon, K. Park, K. Lee, T. Hwang, and T. Chung, Development of a Driver s Intention Determining Algorithm for a Steering System Based Collision Avoidance System, in SAE Technical Paper, no , SAE International, [33] J. Choi and K. Yi, Design and Evaluation of Emergency Driving Support Using Motor Driven Power Steering and Differential Braking on a Virtual Test Track, SAE Int. J. Passeng. Cars - Mech. Syst., vol. 6, no , pp , [34] X. Yang, Prediction of a Vehicle Maximum Forward Speed to Pass Double Lane Change Manoeuvre, International Journal of Vehicle Performance, vol. 1, no. 1,

37 [35] M. Wu, W. Deng, S. Zhang, H. Sun, B. Liu, B. Shang, and S. Qiu, Modeling and Simulation of Intelligent Driving with Trajectory Planning and Tracking, SAE Int. J. Trans. Safety, vol. 2, no , pp. 1 7,

38 CHAPTER 2 DRIVER BEHAVIOR IN COLLISION IMMINENT SITUATIONS 2.1 Introduction The first question asked in the analysis of the application of DBS to heavy vehicles is: do the characteristics of commercial vehicle driver braking behavior indicate that DBS would be effective at assisting these drivers? This question is answered by analyzing a dataset obtained from a prior study of driver braking in emergency scenarios [1]. The main result desired to be drawn from this study is the determination of whether or not commercial vehicle drivers exhibit multi-stage braking. 2.2 Background of Analysis As discussed previously in [2] Prynne and Martin demonstrated that drivers of passenger cars tend to exhibit multi-stage braking in emergency situations. They also stated that this behavior should be expected from all drivers in emergency situations, because humans do not have a proper instinctive braking response in emergency situations. Therefore, one would expect that this behavior would be observed in commercial vehicle driver behavior as well. In order to evaluate whether it is true, that this behavior exists for commercial 18

39 vehicle drivers, a study was conducted of the braking behavior of truck drivers in collision imminent situations [3]. The determination of whether or not this phenomena existed is valuable in that it allows one to evaluate whether Dynamic Brake Support (DBS), a form of brake assist, would be effective in assisting in the avoidance or mitigation of collisions in commercial vehicles. The main question that arises in discussing the implementation of DBS on heavy vehicles is: do commercial vehicle drivers typically apply adequate brake force in a sufficiently short time when presented with collision imminent situations? To answer this question, commercial vehicle drivers behavior during collision imminent scenarios was analyzed. The availability of accident and field research data for crash scenarios involving heavy vehicles is very limited. Therefore, the data used is from a NHTSA National Advanced Driving Simulator (NADS) research project [1] designed to study the effect of reducing heavy truck stopping distance on the number of crashes and their severity. This study was conducted to support the rule making efforts in reducing the Federal Motor Vehicle Safety Standard No. 121 (FMVSS 121) air braked truck stopping distance requirements. Three air braking systems were used; standard S-cam, enhanced S-cam (larger drums and shoes), and an air-actuated disc brake system. A test group of 108 CDL-licensed drivers was split evenly among the simulations using each of the three braking systems. The drivers were presented with four different emergency stopping situations. The four stopping emergency events were right incursion, left incursion, stopped vehicle, and stopping vehicle. These events were simulated on a dry surface and the truck drivers were restricted from steering away from the obstacles by using concrete barriers on the shoulder, parked vehicles, and moving traffic in adjacent lanes. The events were timed such that with a typical S-cam brake system, that met 2004 FMVSS 19

40 121 stopping distance requirements, the driver could bring the truck to a safe stop if he/she pushed the brake pedal to its saturation pressure at the time when the incursion vehicle was first perceived. The results of NHTSA s study [1] showed that the type of braking system had no statistically significant effect on driver behavior prior to braking. Drivers braking efforts measured through pedal forces and reaction times of scenario perceptions were found to be statistically similar. The results validated the theory that; reducing heavy truck stopping distance has strong potential to decrease the number and severity of crashes in situations requiring emergency braking. This is true given that the operators reaction time, control behavior, and their perceptions of available stopping distance remain constant. For a further explanation of the scenarios presented to the subjects, the reader is referred to Appendix A. Table 2.1: Summary of number of collisions with incursion vehicles (Data from [1]). Table 2.1 includes the number of collisions that occurred for each scenario and each brake system type. The right incursion scenario is less likely to result in a collision than the other three scenarios. This may be due to differences in driver s 20

41 perception of these scenarios or driver s having a more instinctive reaction to this case. Comparing these numbers to real world data [4], the right incursion event occurred in 0.5% of all heavy truck multi-vehicle crashes, the left incursion 4.5%, the stopped event 4.2%, and the stopping event 2.9%. Table 2.2 contains mean collision speeds, and the differences in vehicle speeds between the striking vehicle and the struck vehicle, for all the brake types that occurred during the left incursion, stopped event, and stopping event. The right incursion data was not included because there were only three collisions, and a comparison of the mean speeds for this case would be largely meaningless. The left incursion results in a head-on collision; that is why the collision speed is less than the velocity difference. Table 2.2: Mean collision speed and (mean velocity difference) for collisions with simulated vehicles in MPH (Data from [1]). The data from NHTSA s study showed that the ratio of collisions to total number of drivers for the standard S-Cam system is 38.6%, 23.7% for the enhanced S-CAM, and 23.0% for the air disc brake. For the most severe event (the stopped vehicle at 70 mph), the ratio of collisions to the total number of runs for the air disc system is 24.0%, whereas it is 56.0% and 58.6% for the standard S-Cam and enhanced S-Cam 21

42 respectively. Also, drivers using air disc brakes in this event had reduced collision speeds compared with those using the S-Cam and enhanced S-Cam systems. The current study s data analysis did not address the same metrics evaluated during the earlier research. However, the same data from the four crash imminent scenarios was analyzed in the current work to determine driver braking behavior metrics and evaluate the possibility of using DBS to assist in reducing rear-end crashes. The objective of this study was to evaluate heavy truck CDL-licensed drivers braking behavior from the time of danger perception up to maximum braking efforts. The analysis is based on three main concepts: peak brake pressure, application time from brake initialization to when peak brake pressure is reached, and if brake application was performed in multiple stages. In order to ensure that the data being analyzed only consisted of the test subjects response to the presented scenario, each dataset was truncated, so that only the desired time frame was evaluated. Each dataset was specified to start when the scenario began. The end point for the dataset was triggered by one of two events: first, if a driver was involved in a collision; and second, if the driver achieved a complete stop. If neither of these events were triggered, the end of the dataset evaluated was the end of the parent dataset. This typically occurred in situations in which a driver slowed significantly prior to the occurrence of a collision and then allowed the vehicle to coast at a slow speed. In evaluating the datasets to determine whether or not a collision had occurred, it was found that some datasets were incomplete. In these datasets the reported data for the distance between the subject vehicle and scenario vehicles was not recorded correctly. Datasets in which this occurred were removed from the population of datasets used for this analysis, because one could not determine if or when a collision 22

43 had occurred. Also, in some datasets it was discovered that the driver did not apply a brake force exceeding two pounds at any point during the test. Two of such runs were also removed from the population analyzed because the driver did not react to the scenario or reacted in some other way. The resulting population of remaining datasets is presented in Table 2.3. Table 2.3: Population distribution of datasets analyzed (Data from [1]). 2.3 Braking Metrics Various scalar metrics were extracted from each braking profile in order to allow them to be quantified and compared. These metrics are explained in the following sections and results are presented. 23

44 2.3.1 Peak Brake Pressure The brake pressure considered is the control input to the brake system (treadle pressure). The pressure is set using a valve that is attached to the brake pedal and reaches saturation once a given pressure is reached (the vehicle air system pressure). The NADS pneumatic brake pressure model is set to saturate at 100 psi; this level replicates the reservoir pressure for braking in heavy vehicles. Once brake system saturation is reached, any additional force applied to the brake pedal does not increase the system pressure, and in turn does not increase the level of braking. Because of this saturation, it was decided to identify the braking peak based on system pressure, rather than brake pedal force, as it is more directly related to the applied brake torque. Two metrics are computed for every combination of scenario and brake system: peak brake pressure and brake application time. The brake application time is defined as the time between the initialization of braking and when the braking peak is achieved. The peak brake pressure is the brake pressure recorded at the time when a braking peak is identified. The brake system treadle pressure is computed by using the NADS brake model s static curve that relates applied brake pedal force to the treadle pressure that activates the system. It is a static conversion used to determine treadle pressure and when the system has reached saturation. As for the remainder of the analysis in this paper, drivers braking pauses were determined using brake pedal force, since any pause occurs before reaching saturation. The brake force to treadle pressure map is included as Figure Peak Identification The location of the peak is defined as the location of maximum pressure between the times in which the data first crosses the peak start threshold and when it crosses the 24

45 Figure 2.1: Brake force to treadle pressure map for NADS simulation. peak end threshold. The definition of the initialization of braking is the first point in time after the beginning of the scenario when the applied brake force exceeds two pounds. The peak-start and peak-end thresholds are 95% and 90% of the maximum pressure recorded for the desired test, respectively. In the case that the same peak value is recorded at multiple points in time the point that occurs first is selected as the peak. Figure 2.2 is a graphical example of the method of peak identification. It should also be noted that the simulated trailer was heavily loaded to achieve a combined weight of about 80,000 lbs., and with a dry surface the ABS (Anti-Lock Brake System) did not activate for all runs, due to brake torque not being sufficient 25

46 to induce lock-up and therefore ABS activation or wheel lock-up could not be used to determine the occurrence of system saturation. Figure 2.2: Example plot of method of peak identification (Data from [1]). The minimum mean for all datasets in Table 2.4 is 96.1 psi which could indicate that in the majority of tests the driver achieved the brake system saturation pressure at some point during the scenario. To prove this, the data from each test was evaluated to determine if the driver had achieved brake pressure saturation. Table 2.5 contains the results of this analysis. The results of presented in Table 2.4 also 26

47 give the standard deviation for peak system pressure. Being that this distribution is bounded on one side it cannot be normal and therefore this data can only be used to characterize the amount of scatter in the data. This may imply that the data has a reasonably broad distribution about the mean, which also allows for the possibility that a significant number of drivers exceed this mean. From Table 2.5, the column showing the percentage of tests with saturation confirms that a majority of the drivers achieved the brake system saturation pressure. From this table, the minimum percentage is 82% and the overall average percentage for all scenarios is 92%. Based on these results, it is reasonable to conclude that a considerable majority of drivers apply sufficient brake force to achieve maximum system pressure when confronted with a crash imminent situation. Table 2.4: Mean and standard deviation of peak brake pressure for each combination of brake system and scenario (Data from [1]). 27

48 Table 2.5: Percentage of individuals who achieve brake system saturation for each combination of brake system and scenario (Data from [1]) Brake Application Time The brake application time (time from the initialization of braking, to the peak) mean and standard deviation, for each test, were computed for each combination of scenario and brake system. The results from these calculations are presented in Table 2.6. In order to characterize the distribution of brake application time, histograms shown in Figure 2.3 were created using data from all brake systems and scenarios. It appears that the data in the histograms approximates a Rayleigh distribution, or at minimum is not normally distributed. Since the data does not follow a normal distribution the interpretation of the standard deviation is the amount of scatter in the data, and does not indicate a symmetric region of expectation. 28

49 Table 2.6: Mean, standard deviation and p-value of brake application time for each combination of brake system and maneuver (Data from [1]). In looking at the data in Table 2.6, it is also observable that for many of the tests the mean brake application time seems to increase as the brake system is improved. Accordingly, in all scenarios the largest mean brake application time is from the set of drivers with disc brakes. The Kruskal-Wallis test [5] was used to evaluate if a statistically significant difference existed in the mean of brake application time for all drivers between scenarios. The Kruskal-Wallis test is a nonparametric version of ANOVA (Analysis of Variance). The main difference between these two methods is that general ANOVA assumes that the datasets being compared are normally distributed while the Kruskal- Wallis test does not make this assumption. Based on the histograms in Figure 2.3 it is reasonable to assume that the distribution of the data is not normal, and therefore the Kruskal-Wallis test should be used. 29

50 Figure 2.3: Histogram of brake application time for all brake systems and different scenarios (0.15 second bin width) (Data from [1]). The results of statistical analysis in evaluating the effect of the variation in brake system type on brake application time for each scenario is given in the p-value column of Table 2.6. For this study the significance threshold for p-value was selected such that any value below 0.05 disproves the null hypothesis and any value greater than 30

51 0.1 confirms it. The null hypothesis being that the datasets being compared are from distributions with the same mean. The use of the Kruskal-Wallis test generally also includes in the results a box plot to graphically compare test data. Only conclusions using the p-value are presented in this chapter, due to the conclusions from the box plots being open to interpretation. The interpretation of the box plots is left to the reader which are included in Appendix B. The p-values from Table 2.6 indicate that for all scenarios except right incursion the null hypothesis is disproved and therefore the means are statistically different. For the right incursion test the null hypothesis cannot be rejected and the means are statistically similar. It is opined that this difference is due to a difference in driver perception of the severity of the scenario. The Kruskal-Wallis test was also used to determine if there was a statistically significant difference in the brake application time as the test type is varied. The resulting p-value is less than 0.001, which indicates very strongly that the brake application time is scenario dependent. This observation that brake application time is scenario dependent is important because it implies that subjects are applying the brakes in a controlled manner rather than simply slamming on the pedal. Table 2.7: P-values of brake application time comparing each set of brake system type (Data from [1]). 31

52 In evaluating the data from Table 2.6 it was previously noted that in all scenarios the largest mean brake application time is from the set of drivers with disc brakes. In order to further evaluate whether this trend was statistically significant, the Kruskal- Wallis test is used to compare brake application time for all combinations of two different brake systems. The results of this analysis are presented in Table 2.7. Many conclusions can be drawn from the data in this table. First, that there is only a statistically significant difference in brake application time between S-cam and enhanced S-cam systems for the left incursion scenario. Second, in looking at the data comparing enhanced S-cam and disc brake systems all scenarios indicate a statistically significant difference in the data. Finally, three of the four scenarios indicate a statistically significant difference between the S-cam and disc systems. These conclusions are important in that they confirm that drivers with disc brakes often apply the brakes differently than drivers with either type of S-cam system. 2.4 Steering Behavior The four scenarios were designed to discourage drivers from steering. Steering behavior was analyzed to confirm this occurred as desired. In a similar fashion to the braking data, the mean and standard deviation of mean steer angle from scenario start to peak braking for each combination of scenario and brake system were computed. The results from this analysis are presented in Table 2.8. Review of the values for mean and standard deviation yields no discernible trends. The largest mean steer angle has a magnitude of 5.1 degrees and the minimum mean steer angle has a magnitude of 2.3 degrees. This range is on the order of what is typically considered on center behavior [6], that is maintaining a nominally straight path. Based on these values it can be concluded that there was virtually no net steer 32

53 Table 2.8: Mean and standard deviation of magnitude of mean steer angle for each combination of brake system and maneuver (Data from [1]). prior to peak system pressure being achieved and in turn that drivers generally did not try to significantly alter their path of travel in this time frame. 2.5 Multi-Stage Braking The detection of a pause in brake application is used to determine if a driver uses multi-stage braking efforts to reach their peak application level. In this analysis, a brake pause is defined as a period in time, during brake application, in which the derivative of brake force with respect to time (brake force rate) is below a certain level for at least a defined period of time. The threshold used for the brake application rate 33

54 is called the brake pause threshold which is defined as a percentage of the maximum rate during a given test. The minimum duration to be considered a brake pause is called the minimum pause time. The duration of a pause is defined as the time between when the data drops below the pause threshold and when it returns to above the pause threshold. The actual pause durations vary between drivers, so the threshold value and the minimum duration were varied to see if conclusions drawn from this analysis were sensitive to these parameters. Figure 2.4: Example plot of method of pause identification (Data from [1]). 34

55 Figure 2.4 shows a graphical example of the pause detection method. The blue and green curves are the brake pedal force and brake force rate respectively. The brake force rate shown is obtained by taking the numerical derivative of the brake force which is filtered by computing a moving average with a 0.05 second window. The cyan line is the brake pause threshold which is defined as a percentage of the maximum brake rate during the application period. In this specific example the threshold is set at 50%. A threshold value of 25% is also used to compare the sensitivity of pause time analysis to the brake pause threshold. The two broken red lines represent the beginning and end of the detected brake pause. Finally, the magenta line designates the brake force level at which system pressure saturation is achieved; this is generally the peak in tests in which brake saturation is recorded. The plot presented in Figure 2.4 may lead one to believe that all tests that exhibit multi-stage braking have a similar brake pedal force profile. This is not correct. Figure 2.5 displays a limited sampling of brake force profiles all of which exhibit multi-stage behavior. This figure is included in order to show how widely varied the shape of these profiles can be. Table 2.9 displays the percentage of drivers who exhibit multi-stage braking with varying level of brake pause threshold and minimum pause time. In reviewing the data from Table 2.9 it can be observed that, in all cases an increase in the brake pause threshold or decrease in the minimum pause time results in an increase in the percentage of test with multi-stage behavior. It is intuitive that this should occur being that any pause that is detected when using a lower brake pause threshold will be detected with a higher one and that any pause detected with a higher minimum pause time will be detected with a lower one. Though intuitive, this observation is important in that if all pauses dropped below the 25% threshold or were longer than 0.5 seconds these analyses would yield the same results. These observations allow for 35

56 Figure 2.5: Plots of various brake force profiles that exhibit multi-stage behavior (Data from [1]). the inference that the range of brake pause threshold and minimum pause time used in this table are in the range of what is typically expected. Tables 2.1 and 2.2 included in the background section of this Chapter are modified to Tables 2.10 and 2.11, respectively. The only difference is that the columns of each brake type are divided among two groups, those who exhibited multi-staging and those who did not. Table 2.10 shows the number of collisions with and without multi-staging determined using a conservative approach of pause detection, which is using 25% for the brake pause threshold and 0.5 sec for the minimum pause time. As can be observed from Table 2.9, this generates the least number of drivers who exhibited 36

57 Table 2.9: Percentage of tests that exhibit multi-stage braking for each combination of scenario and brake system with varying brake pause threshold and minimum pause time (Data from [1]). multi-staging. With this conservative approach, it can be stated that a DBS system, that detects a braking pause of this level during crash imminent situations, would be helpful to avoid or at least mitigate the severity of 33 of the 109 total simulated collisions. If a brake pause threshold of 50% and a minimum pause time of 0.25 sec were used to determine the presence of multi-staging, the benefits of having a DBS system would be even greater. Moreover, many of the drivers who did not exhibit multi-staging took more time than others to reach brake saturation; assuming that DBS applies the brakes at a high rate a DBS system would also be beneficial in reducing brake application time. Table 2.11 shows that for the stopped scenario, the collision speed is higher with multi-staging. For the left incursion scenario, drivers with disc brakes have 37

58 a higher mean collision speed without multi-staging. Again, this data is based on a conservative approach for detecting multi-stage braking, and a different brake pause detection scheme would provide different results. What is important from this simulation study is that many drivers exhibited multi-stage braking. Also, others were not as aggressive as they could be in applying the brakes and took a longer amount of time to reach maximum braking potential than other drivers. A DBS system would be beneficial in mitigating these simulated collisions. Table 2.10: Summary of number of collisions with incursion vehicles and (number of total tests) - W: with multi-staging; W/O: without multi-staging. Multi-stage braking evaluated with 25% brake pause threshold and 0.5 sec minimum pause time (Data from [1]). 38

59 Table 2.11: Mean collision speed and (mean velocity difference) for collisions with simulated vehicles in MPH - W: with multi-staging; W/O without multi-staging. Multi-stage braking evaluated with 25% brake pause threshold and 0.5 sec minimum pause time (Data from [1]). 2.6 Conclusions In this study the evaluation of braking behavior of commercial vehicle drivers shows that a significant portion of drivers exhibit multi-stage braking behavior similar to that seen in drivers of passenger vehicles [2]. When using the most conservative definition of the presence of multi-staging presented in this document (25% brake pause threshold and 0.5 sec minimum pause time), 33 of the 109 collisions presented involve drivers who exhibit multi-stage braking behavior. This implies that a DBS system capable of detecting multi-staging with these thresholds would be able to assist drivers in avoiding or at minimum mitigating 30% of the simulated collisions in this study. Results from using the Kruskal-Wallis test on brake application time show that a statistically significantly difference in means is present in some cases for both variation in brake system type and scenario. This implies that in many scenarios drivers are 39

60 controlling their brake application rate rather than simply slamming on the brakes. Implementation of DBS in these cases could decrease the brake application time even though multi-staging is not present. The combination of these two arguments implies that the implementation of DBS on commercial vehicles may improve safety, as has been shown for passenger vehicles. Though the results of this study indicate that DBS on heavy vehicles may be able to improve safety, a study tailored to evaluating this system would be necessary in order to achieve conclusive results. 40

61 Chapter 2 References [1] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [2] K. Prynne and P. Martin, Braking Behaviour in Emergencies, in SAE Technical Paper, no , SAE International, [3] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [4] M. Starnes, Large-Truck Crash Causation Study: An Initial Overview. DOT HS , August [5] W. H. Kruskal and W. A. Wallis, Use of ranks in one-criterion variance analysis, Journal of the American statistical Association, vol. 47, no. 260, pp , [6] M. K. Salaani, G. J. Heydinger, and P. A. Grygier, Experimental Steering Feel Performance Measures, in SAE Technical Paper, no , SAE International,

62 CHAPTER 3 FREQUENCY CONTENT BASED ANALYSIS OF DRIVER BRAKING BEHAVIOR 3.1 Introduction When analyzing the braking behavior of passenger or commercial vehicle drivers, the metrics most often discussed are the peak pedal force, or system pressure, and the time from the onset of braking until the peak brake pedal force/ pressure is achieved. This view of events results in braking being perceived as a simple event in which braking is applied over a given time and at a given force, without regard to how this application occurs. Based on the results from Chapter 2 it is know that the shape of the brake application profile is important in analyzing braking behavior. As a result a revised method of braking behavior analysis in proposed and applied in this chapter. 3.2 Motivation Typically, when one thinks of Advanced Driver Assistance Systems (ADAS), systems such as Forward Collision Warning (FCW) and Collision Imminent Braking (CIB) come to mind. In these systems driver assistance is provided based on knowledge about the subject vehicle and surrounding objects. A new class of these systems are being implemented. These systems not only use information on the surrounding 42

63 objects but also use information on the drivers response to an event, to determine if intervention is necessary. As a result of this trend an advanced level of understanding of driver braking behavior is necessary. This chapter presents an alternate method of analyzing driver braking behavior. This method uses a frequency content based approach to study driver braking and allows for the extraction of significantly more data from driver profiles than traditionally would have been done. The use of this method presents opportunities for advancement in the understanding of driver braking behavior and in the evaluation of proposed safety systems which take driver braking behavior as an input Analysis of Panic Events In [1] commercial vehicle driver braking behavior was analyzed and the results from this analysis indicate that a significant percentage of drivers exhibited multi-stage braking in emergency situations. The presence of multi-stage behavior indicates that not only are the brake application time and peak pedal force important but also the shape of the application profile must be considered. Prior research has shown that multi-staging behavior is present and expected in passenger car driver braking behavior as well. In [2] Prynne and Martin express that, two-stage braking is an innate part of the human emergency braking process. This is due to humans not having instinctive reactions to emergency situations and having to make decisions about how much braking should be applied during the process. This indicates that multi-stage braking should be expected in all types of vehicles when drivers are placed in extreme situations. In papers such as that by Prynne and Martin where two-stage braking is 43

64 recognized as being an innate part of the human braking process, a more comprehensive but still fundamental perception of human braking is generally used. This behavior is described similar to a linear characterization of braking, drawn when only peak pedal force and application time are considered, with the exception that a pause at a given time and of a given duration is added. A profile with this characteristic is shown in Figure 3.1. In reviewing Figure 3.1 it can be seen that there are seven parameters necessary to define a brake application profile using this method: F Start is the minimum force required to be deemed brake application. F P ause is the level of force at which the pause occurs. F Max is the maximum brake pedal force in a given application. T o is the time at which braking begins. T 1 is the time from the beginning of braking to when the pause force is reached. T P ause is the duration of the pause. Finally, T a is the total application time. Note, T 2 is also shown in Figure 3.1, but is not an independent parameter because it can be shown algebraically to be defined based on other parameters and is only included for convenience of notation (T 2 = T a T P ause T 1 ). Also, this profile could be simplified to a linear application profile by setting T P ause equal to zero and making the slope for intervals T 1 and T 2 equal. Though able to fit reality more closely than a purely linear model, a large amount of deficiencies remain in this model. One such shortcoming being that, this model is unable to account for instances of decreasing brake pedal force during application. Also, this method of analysis is unable to account for scenarios during which multiple pauses in braking occur. Based on the wide variation in braking profiles scene in reality, both of these situations are expected to be encountered and therefore a method capable of more closely fitting reality should be investigated. 44

65 Figure 3.1: Example plot of a brake force application profile parameterized using a two-stage approach Analysis of Non-Panic Events Similar to the analysis of panic braking events non-panic braking events are often only evaluated based on the peak pressure attained and the duration of the pulse. This view of these events is also very fundamental in that it limits the collected data. If the goal of an advanced braking system is to determine when a driver has exhibited panic like braking behavior an equivalent question that may be asked is to determine when braking behavior is sufficiently different from non-panic behavior. Therefore, this data is of similar value to panic braking data in determining if this is possible and is significantly easier to obtain. By analyzing braking behavior in non-panic situations from a frequency content perspective significantly more data than would typically be drawn can be extracted. This data can help in further quantifying this 45

66 behavior and improving the ability to understand the distinction between panic and non-panic behavior. 3.3 Methodology In order to analyze braking behavior using a frequency content based approach a method based on the Fourier series is used. This method is based on defining a fundamental profile based on the nature of the behavior being modeled and then using the frequency content based method to define the deviation from this curve. Methods for analysis of, and the results from analyzing, both panic and non-panic events with these methods are presented in the following sections Panic Braking Events Figure 3.2: Plot of an example brake application profile plotted with the fundamental slope (Data from [3]). 46

67 The first step in this method is to subtract the linear function that extends from the initialization of braking to the braking peak from the application profile. This linear function is referred to as the fundamental slope and the remaining data is referred to as the profile deviation. An example plot of a brake force profile with the linear application profile is shown in Figure 3.2. While, the profile deviation for the same brake application data is shown in Figure 3.3. The profile deviation is a representation of the drivers deviation from the fundamental slope or linear application profile. Therefore, if a driver followed a linear application profile the values in this dataset would be zero. Since the value of this function is not zero a frequency content fit of this data will allow further information to be gained about the nature of this braking. Figure 3.3: Plot of the profile deviation from the fundamental slope for an example brake application profile (Data from [3]). 47

68 Since the chosen method of analysis is based on the Fourier Series it is important that the derivation of the method starts as such. The expression for the Fourier Series as defined in [4] is given as Equation 3.1. F(t) = A 0 + A n sin(nωt + φ n ) (3.1) n=1 The removal of the fundamental slope forces the value of the profile deviation to be zero at the first and last points. It can thus be decided with the available data that the profile deviation function is half of the fundamental period of a sum of sines function. It can also be concluded that the offset (A o ) and phase (φ n ) must be equal to zero and the angular frequency (ω) must be equal to π/t a where t a is the brake application time. The resulting expression is presented as Equation 3.2. F(t) = A n sin(nπt/t a ) (3.2) n=1 Generally when computing the Fourier Series, the values of A n are computed by using the analytical definition of the desired function. Since in this case the function is known only at discrete points, a different way of computing these coefficients must be used. Also, because the definition of the analytical function is unknown, an infinite number of coefficients cannot be computed. This change is reflected by modifying Equation 3.2 to arrive at Equation 3.3. F(t m ) = n max n=1 A n sin(nπt m /t a ) (3.3) In Equation 3.3, n max is the maximum number of Fourier coefficients used and t m is the time value for a discrete sample. Determination of the maximum number of coefficients that can be computed, for a given dataset, can be done by using the Nyquist Sampling Theorem. This theorem states that the minimum sampling 48

69 frequency of a sinusoid must be two times the frequency which one desires to measure [5]. Equally this can be stated that, given a set of discrete samples from a signal the maximum frequency of sinusoid that can be extracted is equal to half of the sampling frequency. This relationship can be shown in equation form as F max = F sampling /2. Also, a given Fourier coefficient in this method can be related to its frequency by the relationship that F n = n/(2 t a ). By equating these relationships a expression relating the maximum number of Fourier coefficients and the number of samples can be determined as shown in Equation 3.4. n max = F sampling t a (3.4) In viewing Equation 3.4 it is noticeable that the right hand side is equal to the number of samples collected during a given application time. However, since the first and last points are used to determine that the function can be represented by the simplified expression of the Fourier series as expressed in Equation 3.2 (two points are used to determine the phase and frequency), they cannot be used to determine the frequency content of the data. Therefore, the maximum number of Fourier coefficients that can be drawn from a given dataset is the number of data points within a given dataset minus two. Once this has been realized it is possible to determine the value of the Fourier coefficients up to the theoretical maximum. Based on the information in Equation 3.4 it can be shown that the values of t m are defined as designated by the vector T in Equation 3.5. T = [ 0 t 1 t 2 t nmax 1 t nmax t a ] (3.5) The goal of this process is to determine the values of the Fourier coefficients necessary to fit the deviation function. Therefore, Equation 3.3 is modified to reflect 49

70 this goal as shown in Equation 3.6, where D is the profile deviation function defined at discrete points, which is desired to be valid for values of t m from m = 1 : n max. D(t m ) = n max n=1 A n sin(nπt m /t a ) (3.6) The resulting system of equations to be solved is shown as Equation 3.7. ( ) ( ) ( ) πt1 2πt1 nmax πt 1 D(t 1 ) = A 1 sin + A 2 sin + + A nmax sin t a t a t ( ) ( ) ( a ) πt2 2πt2 nmax πt 2 D(t 2 ) = A 1 sin + A 2 sin + + A nmax sin.. ( ) πtnmax D(t nmax ) = A 1 sin + A 2 sin t a t a t a. ( 2πtnmax t a. ) ( ) nmax πt nmax + + A nmax sin t a t a (3.7) This system of equations can be converted to matrix form as shown in Equation 3.8. D(t 1 ) D(t 2 ). D(t nmax ) = sin(πt 1 /t a ) sin(2πt 1 /t a ) sin(n max πt 1 /t a ) sin(πt 2 /t a ) sin(2πt 2 /t a ) sin(n max πt 2 /t a ) sin(πt nmax /t a ) sin(2πt nmax /t a ) sin(n max πt nmax /t a ) A 1 A 2. A nmax (3.8) In Equation 3.8 it can be seen that the matrix of sinusoids is an n max by n max square matrix. Assuming it is non-singular, this results in a system in which the Fourier coefficient vector can be solved for analytically. Though this system can be solved analytically it requires the inversion of a large matrix, and it has been found that other numerical methods are computationally more efficient and produce highly accurate results. Therefore, the coefficients are computed through the use of a least 50

71 squares curve fitting algorithm with the objective function defined as Equation 3.6. The use of this method also allows for solutions to be computed for values of n less than n max. An example of using this data to generate a truncated Fourier series with varying numbers of coefficients from one to n max is shown in Figure 3.4. The first ten Fourier coefficients associated with the fit of this data are given in Table 3.1. Figure 3.4: Example plot of curve fitting the Fourier series to the data while varying the number of Fourier coefficients used (Data from [3]). In looking at the data in Table 3.1 it can be seen that information about the nature of the braking profile at hand can be gained by comparing the relative magnitudes of the Fourier coefficients. The profile presented in Figure 3.4 can be seen to be largely defined by the second Fourier coefficient. This would be expected of profiles that demonstrate two-stage braking similar to what is idealized in Figure 3.1. (along with as a small magnitude value for the first Fourier coefficient) 51

72 Table 3.1: Fourier coefficients A 1 through A 10 for the profile deviation fit shown in Figure 3.4 (Data from [3]). A 1 A 2 A 3 A 4 A A 6 A 7 A 8 A 9 A Non-Panic Braking Events The methodology for evaluating non-panic braking events is similar to that used for the evaluation of panic events with the distinction that there is no fundamental slope to be removed. The reason a fundamental slope is present in the panic braking case is because in a panic situation a majority of drivers apply the brakes until system saturation is reached, or equivalently that drivers apply the brakes such that they start from low pressure and proceed to the system s limit. In non-panic braking this is rarely the case. Drivers often apply the brakes with no intention of reaching full system capacity and therefore the fundamental profile of this application begins and ends at zero. By selecting a fundamental profile of zero this implies that any braking is a profile deviation and can be analyzed without modification. However, in order to allow for braking events to be isolated, thresholds must be set to allow for the beginning and ending of a pulse to be identified. This is because noise in a real measurement system dictates that, not every time the measured brake force leaves zero is the result of a braking event. Therefore a non-panic braking event is described such that it begins anytime brake pedal force exceeds two pounds and ends when pedal force falls below one pound. An example of one such isolated braking event is shown in Figure 3.5 In evaluating the characteristics of these profiles it can be seen that profiles in these cases cannot be negative at any point. This generally implies that the behavior 52

73 Figure 3.5: Plot of an example non-panic braking profile (Data from [3]). Figure 3.6: Plot of an example non-panic braking profile with Fourier series fit (Data from [3]). 53

74 of these profiles is expected to be governed by a half sine function. The brake pulse shown in Figure 3.5 is evaluated using the method as described and the brake pulse profile with the frequency content fit are shown in Figure 3.6. Table 3.2 contains the values of the Fourier coefficients for the fit presented in Figure 3.6. Table 3.2: Fourier coefficients A 1 through A 10 for the profile fit shown in Figure 3.6 (Data from [3]). A 1 A 2 A 3 A 4 A A 6 A 7 A 8 A 9 A Analysis of a Sample Dataset The frequency content method of driver braking behavior analysis has been applied to a dataset consisting of the braking responses of commercial vehicle drivers when presented with collision imminent situations and during general driving. The data set used is the same used for the study presented in [1] and is described as follows. The data presented is originally from a NADS (National Advanced Driving Simulator) study designed to evaluate the influence of decreased stopping distance on the outcomes of simulations in which drivers were placed in collision imminent situations. Four different scenarios, designed to represent a majority of situations seen in the field, were used to evaluate whether the results were situationally dependent and are used to evaluate panic behavior. Also, drivers were allowed to familiarize themselves with the simulator vehicle by being placed in normal driving scenarios, this data was used to analyze non-panic behavior. 54

75 3.4.1 Panic Scenarios and Results The four scenarios are referred to as a left incursion, right incursion, stopping vehicle and stopped vehicle. In a left incursion scenario the driver is presented with a situation in which a vehicle traveling in the opposite direction on a two lane roadway crosses the centerline and proceeds into the drivers path of travel. In a right incursion scenario the driver is placed on a two lane road and a vehicle initially obscured from view pulls into the path of travel of the driver from a perpendicular road to the right. In a stopping vehicle scenario the driver is placed in a situation in which they are following a passenger vehicle on a two lane road and the passenger vehicle driver rapidly decelerates. In a stopped scenario the driver is in the right lane on a four lane road; initially the driver is following another large vehicle, but during the test the lead vehicle rapidly changes lanes, to the left to avoid a stopped vehicle which is now in the subject driver s path of travel. In all situations traffic or obstructions were positioned to prevent drivers from attempting a lane change and therefore limiting the possible response to brake application. Diagrams of all scenarios are included in Appendix A. Further information on these scenarios can be obtained from references [3] or [1]. All braking scenarios in this study including those presented as examples are evaluated using the proposed method. The mean and 95% limits of the first fifteen Fourier coefficients for all tests, in panic scenarios, are presented in Figure 3.7. The limits shown in this plot do not represent the confidence interval but rather represent the symmetric interval about the mean which contains 95% of the data. One important fact that can be drawn from reviewing this data is that the magnitudes of the Fourier coefficients decreases as the frequency increases. This is important in that it indicates that truncation of the datasets, still captures the general character of the deviation function, and has a finite amount of error. Since this trend is true 55

76 for this dataset which contains results from a large number of tests (384), this can be expected for all datasets of a similar type. On the surface this observation may seem to be obvious and expected based on knowledge of Fourier series applications, but the appearance that the values of Fourier coefficients are monotonically decreasing would not be true if the second coefficient(or any other coefficient) dominated behavior (as is expected in two stage braking) or if there was a significant high frequency term associated with some aspect of driver behavior. This observation is important in terms of possible future applications. Figure 3.7: Plot of mean and 95% limits of the Fourier coefficients from panic braking v. frequency in terms of multiples of the fundamental (Data from [3]) Non-Panic Scenarios and Results In the same study from which panic braking data was extracted, drivers were allowed time prior to the panic braking simulations to familiarize themselves with the simulator. This familiarization period allowed drivers to apply the brakes in 56

77 normal situations such as stopping at stop signs and other day to day activities. Each of these profiles was analyzed and braking events meeting the criteria required were extracted and isolated. As previously stated these braking events were defined by starting when brake pedal force exceeds two pounds and ending when brake pedal force falls below one pound. 1,812 total braking events in these sessions meet this criteria. Similar to the panic cases the mean and 95% limits of the first fifteen Fourier coefficients are presented in Figure 3.8. The results in Figure 3.8 show that the braking behavior of drivers in non-panic braking events is largely dominated by the fundamental Fourier coefficient. This indicates that this behavior is largely characterized by a half sine which is as expected. It can also be seen that this behavior is largely dominated by the odd multiples of half sines which implies that the profiles being analyzed contain a significant level of symmetry, which is a reasonable but not necessarily expected conclusion. Figure 3.8: Plot of mean and 95% limits of the Fourier coefficients from non-panic braking v. frequency in terms of multiples of the fundamental (Data from [3]). 57

78 3.5 Conclusions A frequency content based method of analysis has been proposed, derived and used to evaluate a sample dataset. The results generated using this method indicate that when applied to a sample dataset of commercial vehicle drivers in collision imminent situations, meaningful data can be extracted that previously would have been ignored. By using this method opportunities exist to better understand the characteristics of brake application and to develop future applications based on this method. 58

79 Chapter 3 References [1] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [2] K. Prynne and P. Martin, Braking Behaviour in Emergencies, in SAE Technical Paper, no , SAE International, [3] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [4] J. P. Den Hartog, Mechanical Vibrations. New York: Dover Publications Inc., Pages [5] R. S. Figliola and D. E. Beasley, Theory and Design for Mechanical Measurements. New Jersey: Dover Publications Inc., Pages

80 CHAPTER 4 STOCHASTIC BRAKE APPLICATION PROFILE GENERATION 4.1 Introduction In evaluating the behavior of drivers in panic and non-panic situations to determine the response that one would typically expect, the question of whether this behavior could be modeled stochastically arose. This would be desirable in that it would allow candidate systems to be evaluated more broadly than by simply applying data collected from field, or simulator, testing and also could be used to aid in refining algorithms. Using the method of braking behavior analysis explained in Chapter 3 as a starting point this chapter presents the development of such a method. The data from [1] is again used in this chapter as the statistical basis for the model. 4.2 Sampling One component of any stochastic model is the use of statistical sampling to generate pseudo-random samples, using what one knows of the desired behavior. In the case of generating brake force profiles with multi-staging, or generating complex brake pulses, any method of profile generation that can be used requires multiple variables. The method of generating samples used for generating braking profiles combines aspects of 60

81 techniques used for both univariate sampling and multivariate sampling of normally distributed datasets. The resulting method is used to generate datasets with the proper characteristics Generating Pseudo-Random Samples from a Given Univariate Distribution Typically in sampling, once a dataset from which random samples are desired is selected, the next step is to test the fit of various distributions to that dataset in order to determine which is the best fit. This requires that the data closely follows an established distribution and can be fit very closely by such. If the distribution of that dataset does not closely match any of the established distributions, and if the goal is only to generate random samples with the given distribution, this step is not required. A widely used method for sampling any distribution is done through the use of the inverse distribution function [2]. The inverse distribution function is a function that maps the cumulative probability to the value of the data at which that cumulative probability is achieved. This inverse distribution function can be visualized by switching the axes of the Cumulative Distribution Function (CDF). This method is based on the fact that regardless of the distribution of the dataset, the corresponding values of the CDF for the dataset are distributed according to the unit rectangular distribution. Therefore, if random samples from the unit rectangular distribution are obtained, samples from the given distribution can be obtained by reversing the process and mapping through the inverse CDF of the desired distribution. Various algorithms have been developed to generate data according to the unit rectangular distribution making these samples fairly easy to obtain. An example of this for an exponential distribution is given in Figure 4.1 [3]. 61

82 Figure 4.1: Example plot of sampling an exponential distribution using the CDF Generating Pseudo-Random Samples from Any Univariate Distribution The use of this method of sampling, with the inverse distribution functions, generally involves the use of an analytically defined CDF, which is fit to the data. However, by using interpolation it is also possible to generate sample data points numerically through the use of the empirical CDF from the experimental data. This method is limited in that it can only generate data points that are within the range of the original data, but has the value of having face validity in that all data comes directly from experimentation without the use of curve fitting. This method was selected and can be used to generate pseudo-random numbers for any distribution desired. 62

83 4.3 Stochastic Brake Application Time Generation The first step in developing such a model is to determine what variables would be used and to find their distribution function. The first variable desired to be evaluated was the brake application time from onset of braking to peak system pressure, in panic braking, or from the beginning of a pulse to the end of a pulse, in non-panic braking. The brake application time data used for this was drawn from the study of driver braking behavior previously discussed, and the associated familiarization data[4] Panic Braking The brake application time data from all drivers in all scenarios was combined to generate a dataset consisting of values from 384 simulated emergency stopping events. Once this data is consolidated, it is desired to determine what the characteristics of the data are in a statistical sense. In order to do this the Cumulative Distribution Function (CDF) of the data is generated empirically. This is done by sorting the data such that the smallest value is first and the largest value is last to form the x-vector of the CDF. The y-vector is formed by generating a vector that contains all integers from one to the length of the dataset and dividing this vector by the length of the dataset. Once combined there are 384 known points on the empirical CDF. One constraint of the CDF is that it must be zero for a brake application time of zero (because brake application time is limited to being greater than zero), therefore the ordered pair (0,0) is added to the dataset, as it is required to be true. A plot of the empirical CDF generated for this data is shown in Figure 4.2. Based on the sampling method previously presented, this amount of data is sufficient to generate stochastic brake application times for these scenarios. 63

84 Figure 4.2: Plot of empirical CDF of brake application time for panic braking (Data from [1]) Non-Panic Braking The brake pulse length data from the same 1,812 braking events discussed in Chapter 3 are analyzed similar to the brake application times from panic cases. The CDF of this data is computed in order for the general characteristics of the data to be observed in a statistical sense. The CDF for this data is shown in Figure 4.3. Similar to panic braking once this curve is obtained sufficient brake pulse length data is known and samples of this quantity can be generated stochastically Summary of Results This process, as developed, can generate a set of stochastic brake application times, or any other univariate parameter, which can be used to generate basic linear braking profiles or fixed height pulses. Further development is necessary to create more complex braking profiles, which are defined by multiple variables and therefore require 64

85 Figure 4.3: Plot of empirical CDF of brake pulse duration for non-panic braking (Data from [1]). multivariate statistics. This would be favorable in that it would allow multi-staging to be added to panic braking profiles. Furthermore, this would allow for non-panic brake pulses to be generated with significantly more variety of shape. 4.4 Possible Methods of Brake Application and Brake Pulse Profile Generation In the analysis of driver braking behavior previously presented it was found that a significant portion of drivers exhibit multi-stage braking in collision imminent situations [4]. This indicates that many drivers do not apply the brakes in a linear fashion and that, though a model that only allows for the variation in the brake application time may be sufficient for some applications, the addition of stochastic brake profile generation with multi-staging can generate profiles that match reality more closely. 65

86 Similarly, reviewing data from brake pulse profiles shows that these profiles can vary in shape from those resembling square waves and half sines to those that defy simple description. In these cases a more versatile profile generation method can be used to create the wide variety of profiles seen in reality Conventional Methods for Profile Generation As discussed in Chapter 3, there are multiple methods that can be used to quantify and in turn stochastically generate non-panic and panic braking profiles. In the following sections these varying methods are discussed and methods are selected for use going forward Panic Braking Profiles Figure 4.4: Example plot of parameterized multi-staging profile. 66

87 There are multiple ways in which stochastic brake profile generation with multistaging could be implemented. The current state of the art and perception of driver braking behavior implies that generating a brake application profile parameterized such as that shown in Figure 4.4 would be a reasonable and appropriate method to achieve this result. In reviewing this method it can be seen that it is limited in several ways. The first of which being that, the system is only capable of generating brake application profiles that contain a single stage. Also, this method is limited in that it forces the brake application profile to be either increasing or flat and does not allow the brake force to decrease at any point in the profile. The absence of decreasing brake force may be a very important factor in evaluating candidate algorithms because, brake force may decrease at some point in application in an emergency situation and could result in system deactivation (due to perception that a driver desires less braking). This can be summarized to say that this method is limited in that it does not allow for enough variability to represent the breadth of brake application profiles seen in the field. Because one objective of this model is to evaluate commercially available systems, realistic profiles are necessary. This variability is demonstrated in Figure 4.5, which shows various brake application profiles from drivers in collision imminent scenarios from the data used in [4]. Based on the limitations induced by using this conventional method, it was decided that a new method should be developed to generate profiles that accurately represent the characteristics of profiles generated by real drivers. In order to develop this method, further understanding of the deviation of drivers braking behavior from a linear application profile is necessary. 67

88 Figure 4.5: Plot of various braking profiles which demonstrate multi-staging, used to show variability (Data from [1]) Non-Panic Braking Profiles Similar to panic braking, non-panic braking could be characterized in various conventional ways. One can envision that this could be done using a half sine or square wave pulse. A plot of parameterized candidate profiles is shown in Figure 4.6. In viewing these profiles both seem reasonable possibilities, in terms of modeling how drivers could apply the brakes in non-emergency situations. For similar reasons to what is used in evaluating the method for generating panic profiles, including that profiles from real drivers have both increasing and decreasing pedal forces during the pulse, and due to the large variation in braking profiles, there are deficiencies in using this method. Based on these reasons, a frequency content 68

89 Figure 4.6: Example plot of parameterized non-panic brake profiles. based method of profile generation similar to that used for panic brake force profile generation is used for non-panic cases Frequency Content Based Approach to Stochastic Brake Profile Generation Using a method of profile generation based on the process of analysis laid out in Chapter 3, a wide variety of brake profiles can be generated. Similar methods are used for panic and non-panic cases. The methods developed for each type of braking event are presented in the following sections, also example plots of the system output are shown Panic Braking Profiles Based on the frequency content data drawn from analysis as presented in Chapter 3, a method for panic situation brake force profile generation was developed. The 69

90 proposed method created uses two separate force profiles, which are combined to generate the final application profile. The first profile is referred to as the fundamental slope and is defined similar to that used in the method of data extraction. The profile is based on the brake application time, the system saturation force, the braking initialization force and the time at onset of braking. A graphical representation of this profile is shown in Figure 4.7. Figure 4.7: Plot of parameterized fundamental slope. The second profile used is the profile s deviation which is generated such that the sum of sines for the given values is computed and is added to the fundamental slope in the application region (between times T 0 and T 0 + T a ). This profile will be similar in appearance and nature to the profile deviation used in panic braking profile analysis from Chapter 3. 70

91 As was done when extracting frequency content amplitudes from the experimental data, a profile deviation function is added to the fundamental slope and the resulting final profile is derived. The number of coefficients used in this method can vary depending on the number of sine functions desired to be added to the fundamental slope. As previously shown, the fundamental slope is defined by T o, T a, F Start and F Max. The profile deviation is defined by the time variables and an additional vector of variables A, where A := {A 1, A 2, A 3,..., A n } and is the amplitude of the sine functions that make up the profile deviation up to n terms. In order to generate simulated profiles using this method all terms must either be defined to be a fixed value or stochastic samples must be generated. Based on the data available, the method of data collection and the objective of the simulation it is elected to set T o, F Start and F Max as fixed values. T o because evaluation of the effectiveness of an assistance system using this method will be done to determine panic detection, which should be independent of driver reaction time. F Start because its presence only occurs because of the method by which the data was originally processed [4] and therefore it must agree with this value. Finally, F Max is selected because it has been shown that approximately 90% of drivers reach system saturation pressure. Therefore, this variable can be reasonably selected to be fixed at that value. Furthermore, activation is expected to occur at points in time significantly before the peak is reached, and therefore activation should not be influenced significantly by this variable Non-Panic Braking Profiles Non-panic braking profiles can be generated using a method once again similar to that used for data extraction in Chapter 3. This method is similar to that used for panic behavior with the key difference being that the fundamental profile is a constant 71

92 at zero. This is due to the underlying difference in the two application situations. In the panic situation a driver s goal is to go from zero to a high level of pressure. However, in the non-panic case the goal is to go from zero to a certain level, until the desired change in speed is achieved, and then back to zero. This creates a situation in which any force applied is a profile deviation. 4.5 Multivariate Generation of Random Samples Generation of random samples for the purpose of creating stochastic brake profiles, which are more complex than basic linear profiles or fixed height pulses, must be done using multivariate statistics. One obstacle in doing this is that sampling a multivariate distribution except in a few special cases is difficult if not impossible. One multivariate distribution that has been widely researched and has a well defined method of sampling is the multivariate normal distribution. Though this distribution can be sampled with well defined methods, one requirement for using it is that all variables of the datasets which are desired to be sampled, when considered independently, must be normally distributed. This appears to be prohibitive in that the data that is currently desired to sample is not normally distributed. Based on the obstacles involved in sampling an arbitrary multivariate distribution, a sampling method must be devised based on distributions that have established sampling methods. The method that will be used to overcome this involves mapping the original data to a normal distribution in order to use the multivariate normal sampling techniques. The method devised is similar to that used for univariate sampling with the difference being that the multivariate normal distribution, as opposed to the unit rectangular distribution, is used for sampling. In order to do this the original data must be mapped so that all quantities are normally distributed. The first step in this 72

93 method is to use each dataset s empirical CDF to map all datasets to unit rectangular distributions. Next, the data is mapped to corresponding values from a normal distribution using the inverse distribution function of a normal distribution. The Probability Density Function (PDF) of the normal distribution is shown as Equation 4.1 [5]. f(x) = 1 2πσ e (x µ)2 2σ 2 (4.1) Given the desired result of all datasets being normal, we can map to any normal distribution. Therefore for convenience, the distribution to which all parameters would be mapped was chosen as the standard normal distribution. The standard normal distribution is a normal distribution where the mean (µ) is zero and the standard deviation (σ) is one. Once the PDF for this distribution, or any other desired distribution, is obtained, Equation 4.2 can be used to relate the PDF to the CDF [5]. F (X) = X f(x)dx (4.2) The resulting CDF is given as Equation 4.3. Where erf is the error function. F (X) = 1 2 [ ( ) ] X erf (4.3) Given that the goal is to generate a normally distributed dataset from the uniformly distributed dataset, the values of the normal set, X, can be computed for each value from the uniform set, F(X), by numerically solving Equation 4.3. Once this is done the mean vector and covariance matrix for the resulting dataset can be generated. The complete process is summarized in the flow chart in Figure 4.8. The output of this method is the mean vector and covariance matrix of the mapped dataset. These parameters are what is required in order to generate random variables using 73

94 the multivariate normal distribution. Once random samples from the multivariate normal distribution are obtained, it is next necessary to reverse the original mapping process in order to obtain random samples in the native distribution. This process can be envisioned as reversing the initial mapping process and is graphically represented in Figure 4.9. Data in Native Distributions Empirical CDFs Native Data Mapped to Unit Rectangular Distributions Inverse Standard Normal CDF (numerical) Native Data Mapped to Standard Normal Distributions Mean Vector and Covariance Matrix of Mapped Data Figure 4.8: Flow chart of the process of mapping data from the native distribution to the standard normal distribution. Gray boxes represent data mapping operations. Red blocks represent datasets. 4.6 Example Stochastic Braking Profiles Generated Using This Method With all of the building blocks in place this method can now be used to generate braking profiles for both situations. Results for both panic and non-panic cases are presented in the following sections Panic Braking Profiles Using this method pseudo-random parameter sets can be generated and stochastic brake application profiles with frequency content can be generated. An example plot of profiles generated using this method is shown in Figure Ten frequency 74

95 Mean Vector and Covariance Matrix of Mapped Data Pseudo-Random Samples from Multivariate Normal Distribution Standard Normal CDF Pseudo-Random Samples Mapped to Unit Rectangular Distributions Inverse Empirical CDFs Pseudo-Random Samples in Native Distributions Figure 4.9: Flow chart of the process of mapping pseudo-random samples from the multivariate normal distribution to the native distributions. Gray boxes represent data mapping operations. Red blocks represent datasets. coefficients are used in this case because as shown in Chapter 3 these coefficients contain a majority of the information. Also, certain sets of profile deviation coefficients can alter brake application time, since brake application time is a parameter of the profile this change should be limited if not eliminated. By keeping the number of coefficients low fewer samples are rejected Non-Panic Braking Profiles The method presented is also used to generate stochastic non-panic brake pulses. An example plot of brake pulses generated using this method is shown in Figure An increased number of coefficients is used in this case because no set can vary the brake application time. Also non-panic pulse profiles are generally longer and therefore a larger number of coefficients represents a similar maximum frequency. 75

96 Figure 4.10: Example plot of brake application profiles with stochastic frequency content for 8 participants using 10 frequency coefficients. Figure 4.11: Example plot of non-panic brake pulse profiles with stochastic frequency content for 5 participants using 100 frequency coefficients. 76

97 4.7 Conclusions The method of stochastic brake force profile generation presented in this chapter has been shown to be able to generate profiles that match human driver behavior both qualitatively and statistically. Also, the method presented can readily accept new data to expand the model or use a subset of data to restrict the behavior to that present in certain situations. This model and method can be used in the evaluation of candidate algorithms that use brake pedal force as an input to determine the influence of driver behavior on such methods. 77

98 Chapter 4 References [1] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [2] C. Forbes, M. Evans, N. Hastings, and B. Peacock, Statistical Distributions, Fourth Edition. New Jersey: John Wiley and Sons, Inc., Pages [3] L. Devtoye, Non-Uniform Random Variate Generation. New York: Springer- Verlag New York Inc., Chapter 2. [4] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [5] E. O. Doebelin, Engineering experimentation: planning, execution, reporting. McGraw-Hill College,

99 CHAPTER 5 DEVELOPMENT OF A METHOD FOR DETECTING DRIVER BRAKING BEHAVIOR CONSISTENT WITH AN EMERGENCY BRAKING EVENT 5.1 Introduction In order to develop a DBS algorithm for use on heavy vehicles it is necessary to first be able to determine, with some level of confidence, if a driver s response to an event is consistent with panic braking behavior. DBS systems use this determination in conjunction with information on surrounding traffic in order to determine if intervention is necessary. This chapter uses the data from [1] to determine if braking behavior in emergency scenarios can be distinguished from normal driver braking behavior. Various methods have been used in the past to quantify driver behavior and determine if it is consistent with a certain type or pattern of behavior. Generally, these algorithms rely on the use of monitoring a scalar quantity such as the brake pedal speed or pedal force, to determine when they exceed a certain threshold. Brake assist systems for passenger cars, which use brake pedal speed to determine panic by monitoring if an activation threshold has been crossed, are discussed and analyzed in [2], [3], [4] and [5]. A comparison of systems for passenger cars using both brake pedal speed and brake pedal force based systems to determine if braking behavior is 79

100 consistent with a panic response, is presented in [6]. Furthermore, a control system which contains brake pedal rate based brake assist for commercial vehicles is presented in [7] and [8]. These methods produce acceptable results when applied to passenger car and commercial vehicle braking, but further development may lead to better methods. This chapter presents the development of an algorithm for determining if driver braking behavior is consistent with a panic response. A framework is established and variations on that framework are used to test various parameter settings and different activation methods. Results from testing these algorithms on a sample database of driver braking profiles, both from situations expected to result in a panic response and situations in which this response is not expected are presented. This allows the effectiveness of this method to be evaluated and for the evaluation of the probability of false activations. 5.2 Conventional Panic Detection Methodology The development of a methodology of panic detection is based on the use of probability distributions of performance metrics drawn from driver braking behavior, both in panic and non-panic scenarios. Distributions are created for both types of scenarios because the question being asked has a binary outcome. The goal is to determine when a driver s braking behavior is consistent with the driver having panicked during a given test. Being that in the scheme devised it is assumed that there is no partial panic and that all behavior fits into one of these two bins (panic and non-panic); this can be equivalently stated as the goal being to determine when behavior is not consistent with the driver having not panicked. This implies that if either version of this goal can be met, the desired outcome can be achieved. The metrics that are used in this methodology are the peak brake pedal force 80

101 during a given application and the peak brake pedal force rate during an application. The brake pedal force rate is computed by taking the numerical derivative of brake pedal force. This method has an advantage over methods that use different quantities such as brake pedal force and position rate. In this case force sensing is the only driver behavior metric that must be measured whereas in other cases multiple sensors must be used. This allows for sensing to be simplified and in turn the cost of the completed system reduced, this may lead to greater industry acceptance. The following sections present two methods of determining driver intention, which are consistent with the goals as stated above. These goals are; detection of behavior consistent with a panic response and detection of behavior not consistent with a nonpanic response. Though these methods may sound similar the distinction between the two is based on the dataset to which the current application is compared. In the first method behavior is compared with panic behavior while in the second case the current profile is compared with non-panic behavior Conventional Methods of Detecting Driver Intention by Determining When Braking Behavior is Consistent with a Panic Response Given that the prior studies conducted as a part of this work generated a large amount of data on driver behavior, it was decided that the development of a method of detection should begin with that data. The root of this method is the use of the experimental CDF of the data in order to determine the cumulative probability related to the metrics from the current application profile. In order to convert the stream of brake application force and force rate data into scalar metrics that can be compared statistically, a running maximum of these quantities is taken during each test. For both brake pedal force and brake pedal force rate the maximum value for all 81

102 Figure 5.1: Plot of the empirical CDF of maximum brake pedal force from drivers in panic situations (Data from [1]). braking events from both panic and non-panic simulations are logged and the CDF for each case is generated. The CDFs for maximum brake pedal force and brake pedal force rate in panic situations are given in Figures 5.1 and 5.2. In Figure 5.1 it can be seen that a large number of drivers apply lb of brake pedal force. This is the level of system saturation and is specific to the brake system model being used. Therefore, this exact behavior would not be expected with other pneumatic or hydraulic braking systems. One conclusion that can be drawn from this plot is that a majority of drivers do achieve the level of force related to saturation of this specific system. In looking at the data from the CDF for maximum brake pedal force in panic situations presented in Figure 5.1, it can be seen that, approximately 90% of drivers 82

103 Figure 5.2: Plot of the empirical CDF of maximum brake pedal force rate from drivers in panic situations (Data from [1]). achieve brake system saturation, as noted in Chapter 2 (10.68% of drivers have a maximum pedal force below that required for system saturation). This means that if the goal is to detect when driver behavior is consistent with panic behavior using this metric, determining when system saturation pressure is reached would be an effective method. The problem with this method is that detection would only be possible once the driver has already reached the system saturation pressure and therefore the window of time in which assistance was possible has passed. Conversely, when reviewing the CDF for the peak brake pedal force rate presented in Figure 5.2 it can be seen that the CDF is roughly linear. This means that the distribution of this parameter is roughly uniform over the region from 200 lb/s to 1200 lb/s. Further analysis is necessary and comparison of this distribution with 83

104 the distribution of the same metric for non-panic cases is necessary to evaluate the effectiveness of this metric and the levels of activation and false detection at varying thresholds Conventional Methods of Detecting Driver Intention by Determining When Braking Behavior is Consistent with a Non-Panic Response Figure 5.3: Plot of the empirical CDF of maximum brake pedal force from drivers in non-panic situations (Data from [1]). Similar to the panic cases the CDFs of maximum brake pedal force and brake pedal force rate for non-panic cases are shown in Figures 5.3 and 5.4. The CDF of maximum brake pedal force in non-panic situations, Figure 5.3, shows that a majority 84

105 Figure 5.4: Plot of the empirical CDF of maximum brake pedal force rate from drivers in non-panic situations (Data from [1]). of drivers do not reach system saturation pressure in non-panic situations. On further review, it can be seen that in non-panic braking situations 96% of drivers do not reach system saturation force when applying the brakes in non-panic situations. This implies that in terms of distinguishing braking events, after the fact, this method may be very successful. Though, similar to the panic case this method is deficient in terms of detecting panic behavior at low enough force levels to give the system time to act. This can be further analyzed by plotting both CDFs on the same set of axes, as shown in Figure 5.5. Based on the data presented it can be seen that, this metric would be fairly effective if driver braking intention was desired to be determined based on brake pedal force. If a threshold of 52 lb was selected, approximately 98% of panic braking cases 85

106 could be detected with 10% false detections. Also, if the system saturation force were selected 89% of panic cases could be detected with only 3.5% of false detections. As previously stated, this implies that this would be a very reliable means of determining if driver braking was consistent with a panic response. Still, the fact remains that detection could only occur after two thirds of the brake application had occurred. This being true severely limits the possible effectiveness of the system. Figure 5.5: Plot of the empirical CDFs of maximum brake pedal force from drivers in both panic and non-panic situations (Data from [1]). The CDF of maximum brake pedal force rate in non-panic situations, Figure 5.4, shows that in general drivers apply the brakes with a significantly lower brake pedal force rate than drivers do in non-panic situations. This comparison can be 86

107 further explained by plotting the two CDFs on the same set of axes as shown in Figure 5.6. In looking at this plot, the separation between these two distributions becomes more distinct. If it were desired to develop a method of detection based on a threshold of this metric, it can be seen that, if the threshold was set at approximately 200 lb/s, approximately 95% detection would be expected with approximately 15% false detections. Furthermore, if this threshold was set at approximately 300 lb/s, approximately 80% detection would be expected with 6% false detections. Figure 5.6: Plot of the empirical CDFs of maximum brake pedal force rate from drivers in both panic and non-panic situations (Data from [1]). This metric presents promise in panic detection because it is reasonably reliable, and it may detect this behavior earlier than maximum pedal force because it can 87

108 occur earlier in time. Further analysis of this method is necessary to determine if it is a valid means of achieving panic braking detection. 5.3 Panic Detection Methodology Utilizing Multiple Metrics Simultaneously Thus far, all methods of detecting driver braking intention rely on using a single metric to classify the current brake application, such as maximum brake pedal force or maximum brake pedal force rate. Using brake pedal force based detection has shown promise in distinguishing driver braking intention, but does so at a point in time too late to provide assistance. Also, using a brake pedal force rate based method presents deficiencies in that in order to produce high levels of detection, near 95%, high levels of false detection must be accepted, near 15%. Furthermore, in order to have reduced levels of false activation, approximately 6%, detection must be reduced to a lower level, 82%. Though these results seem promising, other methods of detection may be able to improve both sides of this trade-off, by increasing detection while decreasing false detection. One possible method of improving this result is to develop a method of analysis that uses multiple brake profile metrics simultaneously to determine driver intervention. In the following section a candidate method to achieve this result is proposed and results from analyzing driver braking profiles using this method are presented. 88

109 5.3.1 Detection of Driver Intention by Combining Non-Panic Behavior Data for Multiple Metrics The method of detection of panic behavior proposed is designed to approximate the combined value of the percentile when compared with non-panic behavior of a certain set of braking metrics. This method referred to as the Panic Braking Index (PBI) is defined by Equation 5.1. P BI = {1 [(1 P RateNP ) (1 P F orcenp )]} 100 (5.1) Where P RateNP is the value of the CDF of non-panic maximum brake pedal force rate related to the current value of the running maximum of brake force rate. Also where, P F orcenp is the value of the CDF of non-panic maximum brake pedal force related to the current value of the running maximum of brake pedal force. In order to compare this method with the methods previously presented the experimental CDF of the maximum value of the PBI during each test is generated, for both panic and non-panic cases, and is shown in Figure 5.7. In reviewing Figure 5.7, it can be seen that there is significantly more separation between the panic and non-panic profiles than is present in prior methods presented in Figures 5.5 and 5.6. This implies that this method draws a greater distinction between these two behaviors and therefore may be a better candidate for discerning driver intention. In looking at the data more closely, this conclusion can be backed up further by the detection and false detection percentages that can be achieved at various threshold levels. Values of detection percentage of driver panic reactions and percentage of false detections using this method for various PBI threshold levels is presented in Table

110 Figure 5.7: Plot of the empirical CDFs of maximum PBI from drivers in both panic and non-panic situations (Data from [1]). 5.4 Conclusions Review of the data from Table 5.1 shows that, through proper selection of PBI threshold, the percentage of detection relative to percentage of false activations can be significantly improved, being able to achieve 94% panic detection with approximately 5% false detection, compared to approximately 15% for force rate based detection. Though the false detection percentage could be reduced to approximately 4% through using a brake pedal force based detection method, this places the detection threshold very near to the system saturation point, meaning only a very minor improvement could be achieved. Also, when using PBI based detection, if the desired panic detection percentage is reduced to 80% this can be done with a 2.2% false activation 90

111 Table 5.1: Table of PBI panic detection percentage and false detection percentage for varying levels of PBI threshold. PBI Threshold Panic Detection Percentage False Detection Percentage % 31.91% % 21.85% % 13.48% % 8.47% % 7.37% % 5.16% % 3.14% % 2.21% % 1.26% % 0.22% rate, a reduction from 6% with brake force rate only detection. Similar to the previous case, brake force only detection could improve on these numbers. With the exception that, this requires that the detection threshold be set at the saturation point, meaning that panic braking behavior could only be detected once nothing could be done. All results currently available indicate that this method of panic detection is a significant improvement over methods using only one parameter, in terms of detecting whether driver behavior is consistent with a panic response. Further analysis of this method would be necessary to determine the difference in the timing of detection. This comparison would be used in order to determine if detection using this method occurs earlier or later than other proposed methods. This analysis is the subject of continuing development. 5.5 Continuing Work One key element of this method that allows it to evolve significantly as more data is obtained, is that the method currently being considered was developed only using 91

112 driver behavior. This method was developed with and is evaluated using data from a large number of commercial vehicle drivers. This means that this system should generally work well for drivers exhibiting average behavior, but may not be as effective for drivers further from the mean. Being that this method only relies on knowledge of driver data in evaluating panic response this system could be made to be adaptive. The methods presented could be modified such that they generate statistical distributions not based on the behavior of a large number of drivers in limited occurrences, but on a large number of occurrences for the same driver. This would mean that all drivers are evaluated against their own previous behavior. Currently having a false detection is akin to saying that the non-panic behavior of one driver is similar to the panic behavior of another. If the system operated by evaluating a single driver against their past behavior, this would not be true and, false activations may be reduced. Furthermore, the data obtained in familiarization was obtained in a period of time significantly less than what an average commercial vehicle driver would spend behind the wheel in a weeks. This implies that if this system were implemented on a vehicle, after a week it would most likely have more data points in the statistical distributions from one driver than the current method has for all drivers combined. By allowing this method to be tailored to specific drivers, performance is expected to improve, false detections are expected to decrease and driver acceptance should increase. The development of this technique and the further understanding of driver braking behavior, in a general sense, is the subject of ongoing work. 92

113 Chapter 5 References [1] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [2] J. J. Breuer, A. Faulhaber, P. Frank, and S. Gleissner, Real world safety benefits of brake assistance systems, in 20th International Technical Conference on the Enhanced Safety of Vehicles (ESV), [3] S. Kitazawa and Y. Matsuura, Feasibility Study of a Braking Assistant System for Driver Pedal Operation in Emergency Situations, in SAE Technical Paper, no , SAE International, [4] H. Yoshida, T. Sugitani, M. Ohta, J. Kizaki, A. Yamamotoz, and K. Shirai, Development of the Brake Assist System, in SAE Technical Paper, no , SAE International, [5] H. J. Feigel and J. Schonlau, Mechanical Brake Assist - A Potential New Standard Safety Feature, in SAE Technical Paper, no , SAE International, [6] T. Hirose, T. Taniguchi, T. Hatano, K. Takahashi, and N. Tanaka, A Study on the Effect of Brake Assist Systems (BAS), SAE Int. J. Passeng. Cars Mech. Syst., vol. 1, no , pp , [7] D. Zhang, C. Zong, S. Yang, and W. Zhao, Development and Verification of Electronic Braking System ECU Software for Commercial Vehicle, in SAE Technical Paper, no , SAE International, [8] J. Han, Z. Changfu, and Z. Weiqiang, Development of a Control Strategy and HIL Validation of Electronic Braking System for Commercial Vehicle, in SAE Technical Paper, no , SAE International,

114 CHAPTER 6 DEVELOPMENT AND SIMULATION OF A PROTOTYPE DBS ALGORITHM FOR HEAVY VEHICLES 6.1 Introduction With candidate methods of detecting panic like braking behavior identified the next step is to develop a complete DBS algorithm and test its effectiveness. This testing will be more comprehensive than testing using only the panic detection methods in that it will generate data about when detection occurs and how that effects stopping distance and collision outcomes. Also, these simulations include analysis of the effects of setting the TTC threshold at various levels on system activation and performance. The braking profiles from [1] are played back through the DBS algorithm proposed in this study to evaluate their effectiveness. The following chapter presents the DBS algorithm used in this simulation, the simulation model and results from simulations with various sets of detection thresholds. 6.2 Overview of DBS Algorithm used for Simulation Testing The proposed DBS algorithm presented in this study uses two elements to determine if intervention is necessary. The first element is a logical test to determine if the 94

115 Time To Collision (TTC) is below an established threshold. The presence of this threshold is what separates this system from a standard brake assist system, in that sensing of the vehicle s environment is used in conjunction with driver behavior to determine if intervention is necessary. The second element in this system is the panic braking behavior detection segment. This part of the DBS system uses one of the methods identified in Chapter 5 to analyze driver braking behavior to determine if it is consistent with a panic braking response. A diagram of this system is given in Figure 6.1. The dashed line between TTC and the method of panic detection is the subject of continuing work and will be discussed later in this chapter. TTC T T C T hreshold Brake Pedal Force d/dt Panic Braking Behavior Detection AND DBS Activation Figure 6.1: Block diagram of proposed DBS system architecture. 6.3 Selection of TTC Threshold In order to limit false activations a TTC threshold is used such that the system will only activate when a vehicle with which a collision is possible is present and detected. In order to gain insight into when drivers apply the brakes in collision imminent situations the data from the study initially presented in Chapter 2 was reprocessed such that TTC could be extracted. The TTC is calculated using the expression as defined in [2] and is presented in Equation

116 T T C = Range V F ollowingv ehicle V LeadV ehicle : V F ollowingv ehicle V LeadV ehicle > 0 : V F ollowingv ehicle V LeadV ehicle 0 (6.1) The method of collecting data for the lead vehicle in the NADS study is implemented in a manner similar to how radar obtains data in a real system. This implementation is defined such that the subject vehicle only has knowledge of the vehicle directly ahead of it. Also, occasionally the subject vehicle will lose track of the lead vehicle causing intermittent data loss. Though these effects may reduce the effectiveness of the algorithm proposed, they add an element of realism to the simulation by using these streams. Because of this, these effects will be left in the datasets and will therefore be present in further analysis. In order to gain a better understanding of when drivers are applying the brakes in collision imminent situations, TTC values for all drivers, at onset of braking, are recorded. For some of the drivers the onset of braking occurred before a TTC for the lead vehicle had been established. This is primarily true for the right incursion and stopped vehicle cases. It is opined that this is most likely due to some cue about the upcoming event being perceived by the driver before the vehicle was detected by the distance and velocity tracking system. In the right incursion case this is most likely the perception that the approaching vehicle, from the right, is going to cross into the truck s path of travel. In the stopped vehicle case this is most likely the drivers observation of the lead vehicle taking evasive action, implying there is some sort of danger ahead. Diagrams of these two scenarios are included as Figures 6.2 and 6.3. Also, all diagrams for these scenarios along with a written explanation of the cases are included in Appendix A. The mean and standard deviation of TTC at initialization of braking are included 96

117 in Table 6.1. This data is presented both on a scenario dependent basis and with all scenarios combined. In viewing this data it can be seen that the TTC, at initialization of braking, varies significantly depending on the scenario type. Also, as previously stated for the right incursion and stopped scenarios, a large number of tests had an invalid TTC at the onset of braking. Figure 6.2: scenario. Diagram of right incursion Figure 6.3: Diagram of stopped vehicle scenario. 97

118 Table 6.1: Table of mean and standard deviation of TTC at initialization of braking, in seconds, and data for the number of profiles with a valid TTC at initialization of braking(data from [1]). Scenario Type Mean (Seconds) Standard Deviation (Seconds) Number of Profiles with a Valid TTC at Brake Initialization Total Number of Profiles for this Scenario Type Left Incursion Right Incursion Stopping Stopped Overall Figure 6.4: Plot of the empirical CDF of TTC at initialization of braking (Data from [1]). 98

119 In order to gain a better understanding of how this data is distributed, the CDF for the overall set of TTC values at onset of braking is generated and shown in Figure 6.4. The data presented shows the value of the CDF for various values of TTC threshold. This data can be taken to say that for the ratio of braking profiles indicated at a given TTC threshold, that portion of drivers would not have the system activation impeded by the TTC threshold. The remaining portion of drivers may have DBS activation influenced by this parameter. For instance, if four seconds was chosen as the TTC threshold approximately 50% of drivers would fall on each side of this situation. Furthermore, if a threshold of six seconds was chosen, for approximately 85% of drivers DBS activation would not be impeded by this threshold, while the remaining 15% may be influenced. Based on the data currently available, a TTC threshold of six seconds is selected as a candidate threshold for this system. Other thresholds may have different benefits in terms of limiting false activations, but TTC data is currently unavailable for drivers in these situations. The familiarization data used in this study was collected in the absence of simulated traffic and therefore the level of TTC at initialization of braking in normal situations cannot be quantified. Future work, in this area, may include evaluation of this type of data in order to better understand the effects of this threshold. 6.4 Panic Braking Behavior Detection Methodology As discussed in Chapter 5, a wide variety or methods can be used to detect when driver behavior is consistent with panic braking behavior. In Chapter 5, the evaluation of the effectiveness of these methods was limited to comparing detection percentage for various application profiles. This was done without regard for timing of detection or the change in testing outcome had these methods been implemented. The goal 99

120 of the current analysis it to compare the candidate methods based on all metrics of performance. The panic detection method is implemented as shown in Figure 6.1 to test the complete system. The contents of the panic braking behavior detection block in Figure 6.1 are shown in Figure 6.5. Brake Pedal Force P F orcenp Pedal Force Percentile Threshold Panic Braking Index (PBI) Based Panic Detection AND Panic Detection Brake Pedal Force Rate P RateNP Pedal Force Rate Percentile Threshold Figure 6.5: Block diagram of panic detection method. This diagram contains all three methods of detection previously discussed; force based, force rate based and PBI. Through controlling the system thresholds, this methodology allows for these algorithms to be tested independently or in tandem. Because of this all results presented will have three thresholds present to indicate which systems are active, those with non-zero thresholds, and at what level the active systems will have their threshold satisfied. 100

121 6.5 Simulation Diagrams and Code in the MATLAB/ Simulink Environment In order for the vehicle/controller system to be simulated, the co-simulation environment using both TruckSim and MATLAB/Simulink is used. TruckSim is a commercially available vehicle dynamic simulation package that allows for simulations of a wide variety of heavy vehicles to be conducted. This software is designed to enable it to operate in conjunction with the MATLAB/ Simulink environment in order to allow for prototype control and estimation systems to be evaluated. This is done in TruckSim by defining inputs and outputs which are then used to create a S-function block that can be connected to external functions in the Simulink environment. The main Simulink block diagram used to simulate DBS in this study, is presented in Figure 6.6. This level of the Simulink diagram presented consists of many of the major building blocks used in this simulation. Working from left to right the brake force profile generation block uses one of two methods to generate this data. If playback of a previously recorded brake application profile is desired, this data is loaded into MATLAB as a timeseries which can be drawn from to generate the data steam in Simulink. Profiles for brake pedal force playback are generated by creating a timeseries that is initially zero and once a certain point in time is reached the brake application profile desired to be played back is inserted in the time series. The remainder of the time series is a constant equal to the value of brake pedal force at the end of the application profile. This is a reasonable method of completing the brake force profile, because it matches the characteristics of driver behavior observed in the field. This is due to drivers largely maintaining a high level of force once peak force is reached and only releasing once a collision has occurred or to pursue another course of action. 101

122 102 Figure 6.6: Top level of the Simulink diagram used for DBS simulation.

123 Figure 6.7: Diagram of the contents of the Pedal Force Profile Generation Block. If it is desired to use stochastically generated brake application profiles, the frequency coefficients for the desired profile are loaded by MATLAB and the code in this block generates the desired data stream. Figure 6.7 shows the contents of this block. All levels of this section, of the simulation diagram and the associated MATLAB code, are included in Appendix C. The DBS system block contains all of parts of the code that determines if DBS activation is necessary and also includes all of the code necessary to control the brakes after this activation has occurred. The Simulink diagram contained within this block is shown in Figure 6.8. The DBS activation diagram contains the logic as shown in Figure

124 104 Figure 6.8: Diagram of the contents of the DBS system block.

125 The functionality of this diagram is explained from left to right as follows. The DBS pressure control block contains the system that controls the brake system command pressure based on the brake pedal force and the DBS activation state. The output of this block is the driver command pressure, with no DBS, and the composite command pressure, with DBS. The selection between the two is based on the setting of the DBS Enabled variable used in setting up the simulation. The Diagrams and MATLAB code, within these blocks, are included in Appendix C. The PSI to MPa gain is a conversion between the DBS system which operates with pressure in PSI and the TruckSim simulation which uses MPa. The throttle override is a step function used to disable the vehicle throttle when the braking event is started. The outputs of these two blocks are then input to TruckSim. Also, the output from the simulation is saved to the MATLAB workspace in the Simulation Results variable. A plot showing a sample of the driver commanded brake system pressure and the composite brake pressure is shown in Figure Vehicle Model Used for Evaluation of the Prototype DBS System The vehicle model used for simulation of this system is a tractor trailer and therefore two sets of parameters are needed, one for the tractor and one for the trailer. The models used are selected from those developed and validated in [3] and [4]. The tractor model used in a 2006 Volvo 6x4. This model is selected because it has be validated and used extensively for simulations both in the conventional simulation and HIL environments. The trailer selected for this testing comes from the same references stated above and is the trailer loaded to the FMVSS 121 configuration. This trailer configuration 105

126 Figure 6.9: Comparison of brake system pressure with and without DBS (Driver brake profile from [1]). is used because it was designed for use in straight line stopping tests. Therefore, performance in situations similar to those being tested are well documented. As a result, comparison between tests will be based on well defined standards and therefore will be more easily transferable than if another loading configuration had been used. 6.7 Simulation Results and Comparison between Different Combinations of Systems and Thresholds Because of the versatility and variability of the prototype system, various threshold sets must be simulated in order to determine the effectiveness of the system for a wide variety of parameters. This will allow for various panic detection methods to 106

127 be compared and their influence on vehicle stopping performance determined. The following sections present results from analysis of a wide variety of parameter sets aimed at quantifying various aspects of system behavior. All simulations in this study are configured such that they start with an initial velocity of 55 MPH. This is selected because it agrees with a majority of the simulations from which data was collected. Also, 55 MPH is a reasonable speed as it should yield a fair expectation of the systems effectiveness Evaluation of the Influence of TTC on System Effectiveness In order to continue the work on TTC threshold selection and improve upon the available information before setting this threshold, simulations were conducted with the goal being to isolate the influence of this parameter. To achieve this result simulations were conducted with the values of all thresholds, except TTC, set to be near zero. This creates a situation in which virtually any level of brake application, that occurs with a TTC below the TTC threshold, will result in system activation. The sets of thresholds used to achieve this goal are presented in Table 6.2. The first row of Table 6.2 shows that a large TTC threshold above system saturation is used for this analysis. This is used because in reality the computer simulation cannot set the TTC to infinity for cases as defined in Equation 6.1 and therefore a saturation value must be chosen. This saturation value was chosen as the largest integer that can be expressed by a byte (255), because this value is often used for saturation of parameters in systems implemented in the field. By setting a threshold above saturation, this ensures that the TTC is always below this threshold. Therefore, activation should occur for all cases and for virtually any level of braking. This scenario serves as the upper boundary for effectiveness of a brake assist system 107

128 (because a brake assist system does not need a valid TTC) when implemented in these cases, and therefore serves as a benchmark for comparison. Table 6.2: Table of threshold sets used to evaluate the influence of the TTC Threshold on system performance. PBI Threshold Non-Panic Force Percentile Threshold Non-Panic Force Rate Percentile Threshold TTC Threshold Large Above Saturation (L AS ) Large Below Saturation (L BS ) s s s ˆ The value used for a Large Above Saturation TTC Threshold is 256 s ˆ The value used for a Large Below Saturation TTC Threshold is 254 s A second set of simulations with a large TTC below the saturation threshold was evaluated. to the system. This is used to evaluate the influence of adding a TTC threshold The results of this set of simulations show the effect of adding a TTC threshold and show the upper bound of effectiveness for a DBS system (because a DBS system does require a valid TTC). In addition to the simulations previously discussed, additional simulations were conducted to evaluate the influence of the TTC threshold. The parameter sets simulated for this purpose are also presented in Table 6.2. The following sections present the results from simulation, of the variation in key performance metrics as system thresholds are varied. 108

129 System Activation Percentage The first parameter evaluated, in comparing the different settings for the TTC threshold, is the variation in system activation percentage in panic cases based on this parameter. These results are presented in Table 6.3. The results from this analysis show that, as expected, the system with a large TTC threshold above system saturation and low activation thresholds activates for all cases (100%). Furthermore, the results for the scenario of using a large TTC Threshold below saturation show an activation percentage of 90.4%. This implies that by changing from a brake assist system to a DBS system the activation percentage is reduced by 9.6%. This result is different from the percentage of profiles with a valid TTC at brake activation calculated from the data in Table 6.1. This implies that if a TTC is not present, at initialization of braking one may still be present before brake application is completed, and therefore the system may still activate. Table 6.3: Variation in system activation percentage with varying TTC Threshold (Driver brake profiles from [1]). PBI Threshold Non-Panic Force Percentile Threshold Non-Panic Force Rate Percentile Threshold TTC Threshold DBS Activation Percentage L AS 100% L BS 90.4% s 89.3% s 86.7% s 62.0% 109

130 Also, in Table 6.3 results from evaluating various TTC threshold settings is presented. In reviewing these results it can be seen that the activation results for thresholds of eight seconds and six seconds are very similar and near the upper limit. In contrast, using a threshold of four seconds shows a significant decrease in activations. Being that six seconds was identified as a possible candidate for the TTC threshold based on analysis of the TTC at initialization of braking, the result that it is also near the knee point in the system activation percentage curve further supports its selection Change in Brake Application Time The next parameter evaluated in comparing the different TTC thresholds is the brake application time. The brake application time in this study is defined as the time from the initialization of braking to when the peak system pressure is achieved. The results of this analysis are presented in Table 6.4. The results for using a large brake application above saturation show that the upper bound of mean reduction in brake application time, for a brake assistance system as defined in this document, is seconds. By comparing this result with the result from simulation with a large threshold below saturation, it can be seen that by changing from a brake assist system to a DBS system the upper bound of mean reduction in brake application time is decreased by seconds. Further results in this table show the effect on mean reduction of brake application time for variating TTC thresholds. These results indicate that by reducing the TTC threshold from eight seconds to six seconds the mean reduction in brake application time is reduced by seconds. If the threshold is again reduced from six seconds to four seconds, the mean reduction in brake application time decreases again, but by a larger amount of seconds. This inconsistent change in this parameter further 110

131 indicates that similar to other parameters a TTC threshold of six seconds is once again near the knee point of this parameter. Table 6.4: Brake application time results from analysis of variation in TTC threshold (Driver brake profiles from [1]). PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile TTC Mean Brake Application Time With DBS Results Mean Brake Application Time Without DBS Mean Reduction in Brake Application Time when DBS is Added L AS s s s L BS s s s s s s s s s s s s s s s Change in Stopping Distance Further comparison between the results of these simulations can be drawn by looking at the variation in vehicle stopping distance for various parameter sets. The results of this analysis are presented in Table 6.5. In looking at these results is can be seen that the upper limit for reduction in stopping distance when using a brake assist system is ft. When using the large TTC threshold below saturation the mean reduction in stopping distance is reduced by 0.72 ft to 19.52, This number represents the upper limit of reduction in stopping distance for a DBS system and the change between the two tests represents the difference in the mean reduction of stopping distance induced by changing from a brake assist system to a DBS system. 111

132 Similar to all results presented previously, simulation data is presented for multiple TTC thresholds to determine this parameters influence. In reviewing this data it can be seen that when going from a TTC threshold of eight seconds to one of six seconds the mean reduction in stopping distance decreases by 1.28 ft. Similarly, when going from a TTC threshold of six seconds to one of four seconds, the same metric decreases again by 3.73 ft. Once again implying that six seconds is near a knee point for this parameter and is a favorable point for this parameter setting. It is expected that this parameter behaves similarly to the TTC at initialization of braking data presented in Figure 6.4, and therefore only one knee point is expected to occur. Also, further reduction of TTC threshold is therefore expected to significantly impede system effectiveness. Table 6.5: Stopping distance results from analysis of variation in TTC threshold (Driver brake profiles from [1]). PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile TTC Mean Stopping Distance With DBS Results Mean Stopping Distance Without DBS Mean Reduction in Stopping Distance when DBS is Added L AS ft ft ft L BS ft ft ft s ft ft ft s ft ft ft s ft ft ft 112

133 Table 6.6: Change in velocity, at various distances after the initialization of braking, results from analysis of variation in TTC threshold (Driver brake profiles from [1]). 113 PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile Results Mean reduction in vehicle speed (mph) when DBS is enabled at various distances from the onset of braking. TTC 25 ft 50 ft 75 ft 100 ft 125 ft 150 ft 175 ft 200 ft L AS L BS s s s

134 Change in Velocity at Various Distances After Initialization of Braking The last factor used to compare the effects of variation in system performance, as the TTC threshold is varied, is the mean reduction in vehicle velocity at certain distances after the initialization of braking. Comparing these datasets gives an idea of how, if a collision occurs at a certain distance after the driver s response, the outcome of that collision will be varied. These comparisons are drawn for distances after event start ranging from 25 ft to 200 ft using 25 ft intervals. The results from this analysis are presented in Table 6.6. Looking at the results in Table 6.6, the mean velocity reductions follow a similar relationship to other parameters that have been compared. The top row represents the upper bound of this parameter for a brake assist system. The second row represents the upper bound of reduction for a DBS system. All following rows represent the reductions at all points as the TTC threshold is varied. Again, in looking at the data a TTC threshold of six seconds seems to fall near a knee point and therefore will be selected as the threshold value to be used for all simulations going forward. Further understanding of how the change in these parameters influences the velocities at various distances can be gained by looking at the distribution of velocity at these points for all parameter sets presented. Figures 6.10 through 6.14 show the mean velocity and 95% limits for various parameter sets in order to evaluate the system performance at these locations. The 95% limits are defined as the symmetric interval about the mean that contains 95% of the data present. In looking at these plots the solid line connects the means at each location while the error bars represent the 95% limits. In Figure 6.10, it can be seen for a brake assistance system the upper bound of effectiveness produces a result in which all points are clustered very tightly around the mean. This is as would be expected with 114

135 subtle variations most likely occurring due to the driver behavior being different in the time between when the driver applies the brakes and when the system recognizes the application. Next, in Figure 6.11, the variation in velocity at the given points can be seen to be significantly wider than in the previous case. This change is solely due to the system not activating simply when braking is initialized, but rather when that is true along with a TTC below saturation being present. As thresholds are further reduced, Figures 6.12 through 6.14, the error bars for the velocity at these points, with DBS, continue to lengthen and the means begin to visibly shift toward the data for the simulations without DBS. This result is as expected and each plot shows the upper bound of effectiveness of a DBS system with the given TTC threshold. Figure 6.10: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = L AS (256 s) (Driver brake profiles from [1]). 115

136 Figure 6.11: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = L BS (254 s) (Driver brake profiles from [1]). Figure 6.12: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 8 s (Driver brake profiles from [1]). 116

137 Figure 6.13: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 6 s (Driver brake profiles from [1]). Figure 6.14: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with TTC Threshold = 4 s (Driver brake profiles from [1]). 117

138 6.7.2 Comparison of Simulation Results using DBS Systems with Varying Detection Methods In Chapter 5, analysis was done to show that performing panic braking detection using the Panic Braking Index (PBI) showed promise in terms of reducing the number and frequency of false activations. Also, this method shows promise in terms of enabling detection to occur with a high level of confidence at points in time where intervention is still possible. In order to compare these different methods, in terms of their influence on other metrics of performance, simulations were conducted with a series of different parameter sets in order to quantify their effectiveness in improving collision outcomes. The parameter sets used are presented in Table 6.7. These parameter sets are used in the following sections to compare key system performance metrics for the configurations presented. This is done in order to better understand the results associated with the varying panic detection methods presented. For all simulations presented the TTC threshold is set as six seconds because this has been identified as a candidate threshold in various cases and is a realistic value to be used. An example plot of variation in system performance when using one of such systems is shown in Figure System Activation Percentage The first metric used in comparing the methods and thresholds is the system activation percentage. The results of this analysis are presented in Table 6.8. In reviewing this data it is important to remember that, because a TTC threshold of six seconds is used the upper limit of activation percentage for any activation method is equal to 86.7%. In looking at the parameter sets that use PBI the activation percentage generally decreases as the PBI threshold increases. The change in activation follows a similar trend to the data presented in Table 5.1, except the reduction in activation occurs at a decreased rate when a TTC threshold is added. 118

139 Figure 6.15: Example plot of brake system command pressure, velocity and deceleration, with and without DBS, for a single application profile with a PBI threshold of 99.5 and a TTC threshold of 6 seconds (Driver brake profile from [1]). 119

140 Table 6.7: Table of threshold sets used to evaluate the influence of the panic detection method on system performance. PBI Threshold Non-Panic Force Percentile Threshold Non-Panic Force Rate Percentile Threshold TTC Threshold s s s s s s s s For example, when changing from a PBI of 99.5 to 99.8, the change in detection without accounting for the effect of the TTC threshold is 13.8% where in this case the change is reduced to 7.6 percent. This implies that a significant portion of the cases in which activation does not occur are from cases where the TTC in the system was above this TTC threshold for all of the brake application. Also, similar to the analysis presented in Chapter 5, simulation of detection with a brake force rate based method, with a non-panic percentile threshold of 84, presents very similar detection results to using a PBI of Though they have similar detection percentages the PBI based method has a significantly lower false activation probability (5 %) than the force rate based method (16 %) with these thresholds, assuming they are not TTC negated. As well, simulation of a DBS system with a force rate based detection threshold of 94 has a very similar activation percentage to a PBI based system with a threshold of Once again with these parameters the brake force rate based detection has a higher percentage of false activations (6%) than the PBI based system (2%), assuming they also are not TTC negated. Based on this information it can be concluded that, when comparing a PBI based DBS system 120

141 with one that is force rate based a similar level of system activation can be expected with a significantly reduced probability of false activations. Based on the current metrics, the force rate based system produces higher levels of false activation than the other two systems, with a similar detection percentage. Further analysis of other performance metrics is necessary in order to determine if any method presents other significant advantages. Table 6.8: Variation in system activation percentage with varying detection method and parameters (Driver brake profiles from [1]). PBI Threshold Non-Panic Force Percentile Threshold Non-Panic Force Rate Percentile Threshold TTC Threshold DBS Activation Percentage s 85.9% s 84.4% s 83.1% s 75.5% s 83.6% s 77.3% s 84.4% s 79.7% Change in Brake Application Time The next metric used to compare candidate systems is the mean brake application time these results are presented in Table 6.9. In reviewing the data presented, in Table 6.9, the results show that the force rate based systems generally produce a larger mean reduction in brake application time when compared to PBI based systems with 121

142 similar activation percentages. This implies that force rate based systems generally act earlier in time than PBI based systems. Conversely, force based systems tend to produce lower reductions in brake application time for similar activation percentages implying they they respond significantly later in time. The extreme case of this being the force based system with a non-panic force percentile threshold of 96, the large number of system activations present produce only a slight change in mean brake application time. Table 6.9: Brake application time results from analysis of variation in detection method and parameters (Driver brake profiles from [1]). PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile TTC Mean Brake Application Time With DBS Results Mean Brake Application Time Without DBS Mean Reduction in Brake Application Time when DBS is Added s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s s The fact that, force rate based systems produce a larger reduction in brake application time for a similar activation percentage must be weighed against the higher probability of false activation present with these systems. 122 The goal of

143 developing a PBI based DBS system is to balance the effectiveness of a force rate based system with the certainty of activation of a brake force based system. As a results of this it is also reasonable to expect that the performance results would also fall between the two methods Change in Stopping Distance The stopping distance is used next to compare the effectiveness of the various method and thresholds. These results are presented in Table Similar to the results for brake application time, the results presented in Table 6.10 indicate that with the same level of activation, force rate based systems seem to be more effective at reducing stopping distance, and PBI based systems seem to be less effective. Reaffirming that rate based systems tend to act earliest in time with PBI based systems acting next and force based system acting last. Also, the force based systems remain the least effective of the three methods. The reductions in stopping distance may not seem to indicate that the use of DBS presents significant results in terms of reducing collision severity. This is largely because comparing means does not necessarily indicate the desired influence of a DBS system, but is useful in comparing different methods. The goal of a DBS system is not necessarily to shift the mean of vehicle stopping distance, but rather to narrow the range of this metric by bringing outliers of this metric closer to ideal. The shifting of the mean is therefore only a result of this occurring. To illustrate the change in how this distribution appears when DBS is added to the system, the CDFs of stopping distance both without DBS and with all detection methods presented in Table 6.7 are shown in Figure The results presented in Figure 6.16 show how the distribution of stopping distance varies as system parameters are altered. The value of the CDF, at a given stopping 123

144 Table 6.10: Stopping distance results from analysis of variation in detection method and parameters (Driver brake profiles from [1]). PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile TTC Mean Stopping Distance With DBS Results Mean Stopping Distance Without DBS Mean Reduction in Stopping Distance when DBS is Added s 210,7 ft ft 12.8 ft s ft ft 9.1 ft s ft ft 7.6 ft s ft ft 4.7 ft s ft ft 13.4 ft s ft ft 11.1 ft s ft ft 2.9 ft s ft ft 0.1 ft distance, indicates the percentage of drivers who achieve a stopping distance less then or equal to that value for the given thresholds. The data presented in these plots can be extracted to show the percentage of vehicles stopped at a given distance for a given parameter set. For a stopping distance of 220 ft, 60% of drivers without DBS stop in this distance or less. At this same distance, 78% of drivers using a PBI based DBS system, with a threshold of 99.5, are stopped, and 91% of drivers, using a force rate based DBS with a non-panic force rate percentile threshold of 84, have stopped. This type of analysis is more closely related to the systems collision avoidance potential Change in Velocity at Various Distances After Initialization of Braking The final metric used to compare system effectiveness, for the various methods and thresholds presented, is the change in velocity at various distances after the 124

145 Figure 6.16: Plot of the CDFs of stopping distance for the activation methods and threshold sets previously discussed (Driver brake profiles from [1]). initialization of braking. The positions for which the difference in velocity is reported are from 25 to 200 feet in 25 foot intervals. The mean change in velocity for each system at these locations are shown in Table 6.11 Similar to stopping distance the mean of this data does not necessarily characterize the magnitude of this systems influence on stopping performance. In order to further quantify this behavior the mean velocity at these points, along with the 95 percent limits, for all DBS system configurations in comparison with the performance without DBS, is presented in Figures 6.17 through

146 Table 6.11: Change in velocity, at various distances after the initialization of braking, results from analysis of variation in detection method and parameters (Driver brake profiles from [1]). 126 PBI Non-Panic Force Percentile Thresholds Non-Panic Force Rate Percentile Results Mean reduction in vehicle speed (mph) when DBS is enabled at various distances from the onset of braking. TTC 25 ft 50 ft 75 ft 100 ft 125 ft 150 ft 175 ft 200 ft s s s s s s s s

147 Figure 6.17: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 95 (Driver brake profiles from [1]). Figure 6.18: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99 (Driver brake profiles from [1]). 127

148 Figure 6.19: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99.5 (Driver brake profiles from [1]). Figure 6.20: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with PBI based activation with a threshold of 99.8 (Driver brake profiles from [1]). 128

149 Figure 6.21: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force rate based activation with a non-panic percentile threshold of 84 (Driver brake profiles from [1]). Figure 6.22: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force rate based activation with a non-panic percentile threshold of 94 (Driver brake profiles from [1]). 129

150 Figure 6.23: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force based activation with a non-panic percentile threshold of 90 (Driver brake profiles from [1]). Figure 6.24: Plot of mean and 95% limits for velocity at various distances after the initialization of braking with force based activation with a non-panic percentile threshold of 96 (Driver brake profiles from [1]). 130

151 The results from the velocity versus station plots reflect the results presented in Table The results tend to show that the force rate based method of detection produces the largest change in velocity at all points with the PBI based systems being next and the force based systems being the least effective. This result would have to be compared with the lower level of false activations present when using PBI based detection at a similar activation percentage in order for an activation method to be selected. 6.8 DBS System Simulation, Conclusions and Continuing Work The results, of the analysis, presented in this chapter show that using a DBS system with either PBI or force rate based detection presents an opportunity for improving collision outcomes in scenarios in which drivers apply the brakes in a manner that results in a detectable panic response. Both methods and various activation threshold sets can be used to detect this panic behavior. These selections can greatly vary the results associated with the system both in activation percentage and stopping performance. Also, the TTC threshold selected in comparing systems in this study is six seconds. This threshold is selected because it shows promise in not inhibiting activation for a significant portion of the panic events analyzed, while being reasonably small such that false activations can be prevented. The current study does not analyze the influence of this setting on false activations during non-panic braking inputs. The requisite data for this analysis was not present in the current study and data on the TTC for non-panic braking inputs would be necessary to complete this analysis. Further simulation of these systems and other candidate systems should 131

152 allow for further comparison of these different methods. Further opportunities for development of such systems are presented in the remainder of this chapter. One other element, of a DBS system, not addressed in this document is system deactivation once the system has been enabled. Possible methods for doing this include monitoring for a decreasing force rate below a given threshold or determining when the driver has completely released the pedal. Because this development is based on driver behavior after system activation has occurred, data in such situations would need to be collected both in cases where the driver desires activation to continue and in cases when this is not the case. The gathering of this data and the determination of the best method of system deactivation is a topic of further study Proposal of an Expanded Forward Collision Avoidance and Mitigation System with Features of FCW, CIB and DBS In Figure 6.1, at the beginning of this chapter, a dashed line from the TTC to the Panic Braking Behavior Detection Block was shown and discussed as a topic of future work. In all simulations up to this point the thresholds used, in the detection of panic behavior, were set as constants. In the more general sense these detection method thresholds can be seen as a function of the TTC with the current systems only being a special case of this such that the function is a constant. Collision Imminent Braking (CIB), which has been implemented on commercial vehicles, can also be seen as a special case of this system. CIB activates to apply the brakes in situations in which a collision has become unavoidable, this is generally determined by using a TTC threshold. This special case can be visualized as another limit case of this system when a certain collision imminent TTC threshold has been crossed, and the activation threshold is dropped to zero. Having an activation threshold of zero will cause activation to occur once the TTC threshold is crossed. 132

153 This system can be developed by designing the activation threshold as a function, being initially infinite, stepping to a lower value at the TTC threshold presented in the systems tested in this document and converging to zero as the system approaches the collision imminent TTC threshold. The transition between these points can be envisioned as a linear function or something significantly more complex. The determination of the shape of this function and the TTC thresholds are the topic of continuing work. Furthermore, this system would be aided by the collision warning aspect already present in Forward Collision Warning (FCW) systems. The CIB and DBS systems have no ability to assist in cases of an inattentive driver. By combining the effects of FCW, CIB and DBS, a comprehensive system, designed to alert the driver of a situation, assist the driver in improving braking performance should they choose to brake, and finally braking for the driver should they not respond once a collision becomes unavoidable, could be created. 133

154 Chapter 6 References [1] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [2] K. Vogel, A comparison of headway and time to collision as safety indicators, Accident analysis & prevention, vol. 35, no. 3, pp , [3] S. J. Rao, M. K. Salaani, G. J. Heydinger, D. A. Guenther, and F. Barickman, Validation of Real Time Hardware in the Loop Simulation for ESC Testing with a 64 Tractor and Trailer Models, SAE Int. J. Commer. Veh., vol. 6, no , pp , [4] S. J. Rao, Development of a Hardware in the Loop Simulation System for Heavy Truck ESC Evaluation and Trailer Parameter and State Estimation. PhD thesis, The Ohio State University,

155 CHAPTER 7 CONTRIBUTIONS TO THE ENGINEERING COMMUNITY AND CONTINUING WORK 7.1 Contributions Based on the work presented in this document, the following contributions to the engineering community have been made. 1. An analysis of commercial vehicle driver braking behavior in collision imminent situations has been presented. This analysis contributes a large amount of information on driver braking behavior including statistics on the percentage of drivers who achieve brake system saturation and driver brake application time from onset of braking until the peak system pressure is achieved. Further analysis of this behavior has shown that commercial vehicle drivers exhibit multi-stage braking similar to that seen in passenger car drivers. This conclusion is controversial in that it does not agree with the beliefs of some portions of the industry. This belief is best characterized by the notion that, commercial vehicle drivers are professional drivers and therefore achieve peak system pressure faster than passenger car drivers. Also, due to the differences in the characteristics of pneumatic and hydraulic brake systems, driver behavior in vehicles with pneumatic systems is different than behavior in similar situations with hydraulic systems. 135 The

156 presence of multi-staging indicates that some safety systems, namely Brake Assist (BA) and Dynamic Brake Support (DBS), could be effective in improving performance in these situations. 2. A novel method of analyzing driver braking behavior using a frequency content based approach has been created and applied. This method of driver braking behavior analysis extracts significantly more data from braking profiles than would traditionally be done. This type of analysis allows for both panic and non-panic braking events to be better understood and characterized. The results of this analysis indicate that though a large number of frequency coefficients can be generated to characterize driver braking, the magnitudes of these coefficients decreases as frequency increases. This implies that truncation of these datasets can be done without losing information about the overall nature of the dataset. Also, the difference between the magnitude of even and odd frequency coefficients for non-panic brake pulses indicates that these pulses are reasonably symmetric in nature. 3. A new method of stochastic driver brake force profile generation has been developed. The method is capable of being used to generate profiles both for panic and nonpanic braking events. The development of this method is based on the method of frequency content based analysis previously established. This method uses multivariate statistics and data mapping in order to generate complex datasets with the correct statistical characteristics. The use of this method in evaluating candidate methods of driver assistance can aid in system development because with a small sampling of driver behavior, less than five hundred profiles, 136

157 thousands of profiles can be generated. This allows for the robustness of the method to be further tested. 4. Development of a new method of driver panic braking behavior detection and comparison with other methods. Based on the new information on commercial vehicle driver braking behavior discussed earlier; it was decided to perform an initial analysis to determine if the characteristics of commercial vehicle driver braking behavior are conducive to panic detection. This method of detection is one of the fundamental building blocks in development of a DBS or BA system. Methods of detection of the type currently being implemented and researched show some positive results but contained some deficiencies. The main deficiency being that in order to obtain high levels of detection one would be required to accept a large number of false activations, or conversely to limit false activations one would have to accept lower levels of detection. In an attempt to shift this balance, such that acceptable levels of detection can occur with lower levels of false detection, a revised method of panic detection was presented and analyzed. This method succeeds in achieving this goal and results in a significantly reduced amount of false detections or a significantly increased level of detection depending on in which direction the results are pushed via the detection threshold. 5. Development and simulation of a DBS algorithm tuned for commercial vehicle driver behavior Using the methods of panic braking behavior detection previously discussed a prototype DBS system was developed and simulated using driver braking inputs collected from the previously presented study as the system input. This system was evaluated using multiple detection methods and system thresholds to gain 137

158 a comprehensive picture of the possible effectiveness in terms of reduced brake application time, velocity, and stopping distance. The results of this study show that this system has a strong potential to improve safety and further development of this method and technology should be pursued. 7.2 Continuing Work The narrative of the work presented spans from the initial evaluation of the possibility of using such a system on commercial vehicles; to the realization that driver behavior indicates that such a system could improve stopping performance in situations in which panic braking is necessary; through the evaluation and comparison of candidate methods. This analysis lays much of the ground work necessary to justify further evaluation of this technology on commercial vehicles. The current state of the work is by no means considered a conclusion but simply a significant landmark on the way to fully understanding the role these systems can play in the greater realm of active safety on commercial vehicles. Because of this, significant amounts of continuing work are presented, as they will allow more pieces of this puzzle to be understood Evaluation of DBS using HIL Simulation to Evaluate its Effectiveness in Conjunction with Other Safety Systems By implementing DBS in a HIL environment the effectiveness of this system, in the presence of various other safety systems, can be evaluated. The commercially available systems that can be simulated in conjunction with DBS include, ABS, ESC and CIB. Also, by simulating in the HIL environment, the physics of a real brake system will be included in order to evaluate the influence of the pneumatic system dynamics on system performance. The HIL system currently in use by VRTC is shown in Figure

159 Figure 7.1: Photograph of NHTSA s heavy vehicle pneumatic HIL system. In order to allow for this to be simulated a means of controlling brake pedal force, to emulate previously recorded driver brake application profiles and stochastically 139

160 generated profiles, must be created. To achieve this goal a brake pedal force controller has been fabricated and implemented on the HIL system. A photograph of this system is shown in Figure 7.2. Figure 7.2: Photograph of the brake pedal force controller attached to the pneumatic HIL system. The results of simulation, using the HIL, are expected to be used to validate the simulation results presented from the TruckSim/ MATLAB environment. Several steps toward this testing have occurred, such as development of the brake pedal force controller and implementation of a supplemental pneumatic system to allow for DBS 140

161 and CIB system to be implemented in the HIL environment. The remaining steps consist of, implementation of the offline simulation code in the HIL environment and validation of the supplemental pneumatic system and the brake pedal force controller Use of Stochastic Brake Pedal Force Generation in Evaluation of Various DBS Activation Methods and Thresholds The current comparison of DBS systems is limited to that based on playback of driver braking profiles from simulation. Further insight may be gained by implementing stochastic brake application profile generation to evaluate these methods. Also, simulation using this technique would be helpful in developing a method of recording driver braking behavior in order to tune the system s activation profiles. By allowing the system to evolve over time the effect of this evolution on the results of simulation could be evaluated. This simulation would still be performed using behavior from a large number of drivers. Though, by selecting subsets of the greater population with a certain type of behavior, and using that behavior to generate the stochastic model, the evolution of the systems for groups of drivers with more or less aggressive braking inputs could be evaluated and compared to results for the complete population. The current limitation on performing these simulations is computation time. The real world time necessary to perform a braking simulation, with the current hardware, is approximately 45 seconds. Implying that, to run the 384 events presented in this study the average computation time is approximately 5 hours. Further, this implies that to run the 2,000 stochastic events, 25 hours would be necessary not including post-processing time. 141

162 BIBLIOGRAPHY [1] NHTSAs National Center for Statistics and Analysis, Traffic Safety Facts, 2012 Data, Large Trucks. DOT HS , May [2] P. M. Knoll, B.-J. Schaefer, H. Guettler, M. Bunse, and R. Kallenbach, Predictive Safety Systems - Steps Towards Collision Mitigation, in SAE Technical Paper, no , SAE International, [3] E. N. Mazzae, F. Barickman, G. H. S. Baldwin, and G. Forkenbrock, Driver Crash Avoidance Behavior with ABS in an Intersection Incursion Scenario on Dry Versus Wet Pavement, in SAE Technical Paper, no , SAE International, [4] L. Evans, ABS and Relative Crash Risk Under Different Roadway, Weather, and Other Conditions, in SAE Technical Paper, no , SAE International, [5] R. Emig, H. Goebels, and H. J. Schramm, Antilock Braking Systems (ABS) for Commercial Vehicles - Status 1990 and Future Prospects, in SAE Technical Paper, no , SAE International, [6] S. B. Zagorski and R. L. Hoover, Comparison of ABS Configurations and Their Effects on Stopping Performance and Stability for a Class 8 Straight-Truck, in SAE Technical Paper, no , SAE International, [7] M. L. Shurtz, G. J. Heydinger, D. A. Guenther, and S. B. Zagorski, Effects of ABS Controller Parameters on Heavy Truck Model Braking Performance, in SAE Technical Paper, no , SAE International, [8] H. Yoshida, T. Sugitani, M. Ohta, J. Kizaki, A. Yamamotoz, and K. Shirai, Development of the Brake Assist System, in SAE Technical Paper, no , SAE International, [9] K. Prynne and P. Martin, Braking Behaviour in Emergencies, in SAE Technical Paper, no , SAE International,

163 [10] S. Chakraborty, T. A. Gee, and D. Smedley, Advanced Collision Avoidance Demonstration for Heavy-Duty Vehicles, in SAE Technical Paper, no , SAE International, [11] J. Jansson, J. Johansson, and F. Gustafsson, Decision Making for Collision Avoidance Systems, in SAE Technical Paper, no , SAE International, [12] P. Ruecker, Crash Tests with Automatic Pre-Crash Braking Cars, in SAE Technical Paper, no , SAE International, [13] M. Egelhaaf and P. Rücker, Incidence of Frontal Impact Accidents and Crash Testing of Cars Equipped with Collision Imminent Braking Systems, in SAE Technical Paper, no , SAE International, [14] M. Lindman and E. Tivesten, A Method for Estimating the Benefit of Autonomous Braking Systems Using Traffic Accident Data, in SAE Technical Paper, no , SAE International, [15] H. Enomoto, K. Akiyama, and H. Okuyama, Advanced Safety Technologies for Large Trucks, in SAE Technical Paper, no , SAE International, [16] J. Woodrooffe, D. Blower, C. A. C. Flannagan, S. E. Bogard, P. A. Green, and S. Bao, Automated Control and Brake Strategies for Future Crash Avoidance Systems - Potential Benefits, in SAE Technical Paper, no , SAE International, [17] J. Woodrooffe, D. Blower, C. A. C. Flannagan, S. E. Bogard, and S. Bao, Effectiveness of a Current Commercial Vehicle Forward Collision Avoidance and Mitigation Systems, in SAE Technical Paper, no , SAE International, [18] H. J. Feigel and J. Schonlau, Mechanical Brake Assist - A Potential New Standard Safety Feature, in SAE Technical Paper, no , SAE International, [19] T. Hirose, T. Taniguchi, T. Hatano, K. Takahashi, and N. Tanaka, A Study on the Effect of Brake Assist Systems (BAS), SAE Int. J. Passeng. Cars Mech. Syst., vol. 1, no , pp , [20] S. Kitazawa and Y. Matsuura, Feasibility Study of a Braking Assistant System for Driver Pedal Operation in Emergency Situations, in SAE Technical Paper, no , SAE International, [21] J. J. Breuer, A. Faulhaber, P. Frank, and S. Gleissner, Real world safety benefits of brake assistance systems, in 20th International Technical Conference on the Enhanced Safety of Vehicles (ESV),

164 [22] D. Zhang, C. Zong, S. Yang, and W. Zhao, Development and Verification of Electronic Braking System ECU Software for Commercial Vehicle, in SAE Technical Paper, no , SAE International, [23] J. Han, Z. Changfu, and Z. Weiqiang, Development of a Control Strategy and HIL Validation of Electronic Braking System for Commercial Vehicle, in SAE Technical Paper, no , SAE International, [24] P. M. Knoll, Predictive Safety Systems: Convenience - Collision Mitigation - Collision Avoidance, in SAE Technical Paper, no , SAE International, [25] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [26] G. F. Fowler, R. E. Larson, and L. A. Wojcik, Driver Crash Avoidance Behavior: Analysis of Experimental Data Collected in NHTSAs Vehicle Antilock Brake System (ABS) Research Program, in SAE Technical Paper, no , SAE International, [27] W. G. Najm and D. L. Smith, Modeling Driver Response to Lead Vehicle Decelerating, in SAE Technical Paper, no , SAE International, [28] K. Vogel, A comparison of headway and time to collision as safety indicators, Accident analysis & prevention, vol. 35, no. 3, pp , [29] K. D. Kusano and H. Gabler, Method for Estimating Time to Collision at Braking in Real-World, Lead Vehicle Stopped Rear-End Crashes for Use in Pre- Crash System Design, SAE Int. J. Passeng. Cars Mech. Syst., vol. 4, no , pp , [30] L.-k. Chen, C.-c. Dai, and M.-f. Luo, Investigation of a driver-oriented adaptive cruise control system, International Journal of Vehicle Design, vol. 66, no. 1, pp , [31] D. Lechner and G. Malaterre, Emergency Manuever Experimentation Using a Driving Simulator, in SAE Technical Paper, no , SAE International, [32] T. Hong, J. Kwon, K. Park, K. Lee, T. Hwang, and T. Chung, Development of a Driver s Intention Determining Algorithm for a Steering System Based Collision Avoidance System, in SAE Technical Paper, no , SAE International,

165 [33] J. Choi and K. Yi, Design and Evaluation of Emergency Driving Support Using Motor Driven Power Steering and Differential Braking on a Virtual Test Track, SAE Int. J. Passeng. Cars - Mech. Syst., vol. 6, no , pp , [34] X. Yang, Prediction of a Vehicle Maximum Forward Speed to Pass Double Lane Change Manoeuvre, International Journal of Vehicle Performance, vol. 1, no. 1, [35] M. Wu, W. Deng, S. Zhang, H. Sun, B. Liu, B. Shang, and S. Qiu, Modeling and Simulation of Intelligent Driving with Trajectory Planning and Tracking, SAE Int. J. Trans. Safety, vol. 2, no , pp. 1 7, [36] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [37] M. Starnes, Large-Truck Crash Causation Study: An Initial Overview. DOT HS , August [38] W. H. Kruskal and W. A. Wallis, Use of ranks in one-criterion variance analysis, Journal of the American statistical Association, vol. 47, no. 260, pp , [39] M. K. Salaani, G. J. Heydinger, and P. A. Grygier, Experimental Steering Feel Performance Measures, in SAE Technical Paper, no , SAE International, [40] J. P. Den Hartog, Mechanical Vibrations. New York: Dover Publications Inc., Pages [41] R. S. Figliola and D. E. Beasley, Theory and Design for Mechanical Measurements. New Jersey: Dover Publications Inc., Pages [42] C. Forbes, M. Evans, N. Hastings, and B. Peacock, Statistical Distributions, Fourth Edition. New Jersey: John Wiley and Sons, Inc., Pages [43] L. Devtoye, Non-Uniform Random Variate Generation. New York: Springer- Verlag New York Inc., Chapter 2. [44] E. O. Doebelin, Engineering experimentation: planning, execution, reporting. McGraw-Hill College, [45] S. J. Rao, M. K. Salaani, G. J. Heydinger, D. A. Guenther, and F. Barickman, Validation of Real Time Hardware in the Loop Simulation for ESC Testing with a 64 Tractor and Trailer Models, SAE Int. J. Commer. Veh., vol. 6, no , pp ,

166 [46] S. J. Rao, Development of a Hardware in the Loop Simulation System for Heavy Truck ESC Evaluation and Trailer Parameter and State Estimation. PhD thesis, The Ohio State University,

167 APPENDIX A EXPLANATION OF SCENARIOS PRESENTED TO DRIVERS AND VEHICLE CONFIGURATIONS FOR NADS SIMULATIONS This section on the scenario design for the NADS simulations has been included in prior publications, ([1] and [2]) and is included to serve as a reference to the reader. The effectiveness of improved brakes on heavy trucks is examined using three different brake system conditions and four simulator scenarios. The three different brake configurations were: ˆ Standard truck where S-cam brakes were used on all wheels. ˆ Enhanced S-cam truck where only the steer axle was equipped with a higher capacity version of an S-cam brake. ˆ Air disc truck where all the wheels of the tractor were equipped with air disc brakes. The simulator scenarios were primarily based on those used in previous NHTSA Electronic Stability Control (ESC) research [3]. All simulated roads were built with a shoulder with different traction, vibration, and audio characteristics than the on-road pavement. This is to realistically simulate the environment that occurs when some 147

168 of a vehicle s tires depart the roadway. The lanes were 12 feet (3.7 m) wide, there were 1.9 feet (0.58 m) of road between the white line (designating the outboard edge of the lane) and the shoulder, the shoulder was 11.5 feet (3.51 m) wide. Beyond the shoulder, there was an additional 75 feet (23 m) of drivable terrain (see Figure B1). The scenarios took place on dry pavement. The virtual environment reflected conditions consistent with the pavement. In particular, the scene was clear and the pavement appeared dry. The study used the NADS heavy truck cab and dynamics model [2]. A typical 18- wheel tractor-trailer combination was selected with a gross weight of 73,100 pounds (33,200 kg). Stopping distance was reduced by 17% and 30% when the standard S- cam brake system was replaced by the enhanced S-cam and disc systems respectively. Truck drivers were recruited from local Iowa trucking companies as well as through radio and newspapers ads targeted at all truck drivers in the area. Participants consisted of drivers who held a valid Commercial Drivers License (CDL) and were between the ages of 22 and 55 (current statistics show that approximately 75% of all drivers involved in heavy truck crashes are between the ages of 22 and 55 and drove, on average, 2000 miles during the last 3 months). This ensured that participants were actively driving heavy trucks. The population of commercial vehicle drivers is comprised of mostly males, but no attempt was made to balance by gender. Participant pay, in this experiment, was comparable with a professional truck drivers hourly wage of $30 per hour plus incentive pay. A repeated measures experiment design in which participants experienced multiple scenarios was used. Independent variables included brake system (3 levels: standard S-cam, enhanced S-cam, and air disc brakes) and event order (4 events were used, but only 3 events were fully randomized, giving 6 levels; the fourth event was always last). A single age group was used (22-55). This design resulted in 18 experimental 148

169 Figure A.1: Road geometry. cells. To allow 6 repetitions of each event order per brake condition, 108 participants who would successfully complete all 4 events were needed. This recruiting goal was met. The principal measure for this study was whether the driver crashed or not. Secondary measures consisted of collision speed, stopping distance, reaction time to event start, and average deceleration. Other behaviors were tabulated such as if the driver braked, steered, and/or accelerated. 149

170 A.1 Scenario Design To understand the effectiveness of heavy truck improved brakes, scenarios were designed to emulate real world situations where heavy truck crashes are occurring. Dry asphalt pavement conditions were simulated. A total of four scenarios containing situations conducive to emergency braking were used. Events were presented to each participant as individual drives. Each participant drove all the scenarios. Each scenario was approximately five minutes in length and ended immediately after presentation of a conflict event. The scenarios were designed to have consistent entry speed (maintained through monetary incentives) for all participants and no downshifting during the event itself. They were also designed such that the driver could stop without hitting the target vehicle, if the brakes were applied immediately. The scenarios conflict events were: Right Incursion The goal of this event was to force the driver to apply brakes to avoid colliding with oncoming traffic. A vehicle that pulls out of a hidden driveway attached to a roadside farmhouse combined with carefully timed oncoming traffic created the conditions for such a maneuver (Figure A.3). The driver approached a driveway that can hide a vehicle. The driver was motivated via monetary incentives to maintain the speed limit of 55 mph (89 kph). Parked vehicles on both shoulders prevented the driver from steering to avoid the oncoming traffic. When the driver was four seconds from arriving at the driveway location, the hidden parked vehicle pulled out from the right and stopped, blocking the right lane. Drivers who could not stop within the available distance would collide with the white incursion vehicle, the green oncoming car, the gray oncoming car, or the parked truck on the left shoulder. 150

171 Left Incursion The goal of this event was to force the driver to react to an incursion from the left and to brake suddenly while traveling at highway speed. The driver was on a two-lane rural highway crossing a heavily wooded area with frequent oncoming traffic (Figure A.2). The posted speed limit was 55 mph (89 kph) and the driver was motivated via monetary incentives to maintain speed. There were several parked vehicles on both shoulders. As the driver approached the location of the event, one of the oncoming vehicles was tasked to arrive at the event location at a fixed relative position to the driver. Oncoming traffic approached a parked vehicle on the shoulder opposite to the driver s side. That parked vehicle began moving and cut off the oncoming traffic which was then forced to steer into the driver s lane. The oncoming traffic would enter the driver s lane at a fixed time-distance, 8 seconds away from the driver. Concrete barriers were placed on the right side so that the driver would not steer to the shoulder. If the driver could not stop within the available distance, the driver would collide with the oncoming red SUV, oncoming traffic in the left lane, or the concrete barriers. Stopped Vehicle The goal of this event was to force the driver to react to an obscured stopped vehicle on the highway. The driver was on a 4-lane rural highway traveling at the posted speed limit of 70 mph (110 kph) (Figure A.5). There was a steady stream of traffic in the adjacent lane as well in the oncoming lanes. Once the driver achieved the posted speed limit, a delivery truck sped past, made a right lane change into the driver s lane, and became the lead vehicle as well as the obscuring vehicle. The lead vehicle maintained a distance of 400 ft (122 m) in front of the driver. When the participant was 610 ft (186 m) from a stopped 151

172 vehicle, the lead vehicle made a lane change into the stream of traffic in the adjacent lane revealing the stopped vehicle. The driver could collide with the stopped vehicle, traffic traveling in the same direction in the adjacent lane, or oncoming traffic in the far lanes. Because this was the most severe stop, it was always the last scenario for each driver. Stopping Vehicle The goal of this event was to force the driver to react to an abruptly stopping lead vehicle while traveling at 55 mph (89 kph). There was a continuous flow of oncoming traffic throughout the event and there were barricades and construction vehicles parked along the right side of the road. These barricades and parked vehicles constrained the driver from steering off-road during the braking event (Figure A.4). The driver was on a two-lane rural highway crossing a heavily wooded area with frequent oncoming traffic. The posted speed limit was 55 mph. As the driver was moving along, a vehicle approached the truck from behind. The following vehicle then made a lane change and overtook the truck. It entered the driver s lane and maintained a distance of 132 ft (40 m) for approximately 2100 ft (640 m) before it decelerated at the rate of 0.75 g to a complete stop. The driver was discouraged from steering via construction barriers on the edge of the driver s lane and oncoming traffic in the adjacent lane. A collision could happen with the stopping green lead vehicle, oncoming traffic, or the concrete barriers. 152

173 Figure A.2: Left incursion scenario diagram. 153

174 Figure A.3: Right incursion scenario diagram. 154

175 Figure A.4: Stopping vehicle scenario diagram. 155

176 Figure A.5: Stopped vehicle scenario diagram. 156

177 Appendix A References [1] J. L. Every, M. K. Salaani, F. S. Barickman, D. H. Elsasser, D. A. Guenther, G. J. Heydinger, and S. J. Rao, Braking Behavior of Truck Drivers in Crash Imminent Scenarios, SAE Int. J. Commer. Veh., vol. 7, no , pp , [2] M. K. Salaani, G. J. Heydinger, P. A. Grygier, C. Schwarz, and T. Brown, Study of Heavy Truck S-Cam, Enhanced S-Cam, and Air Disc Brake Models Using NADS. DOT HS , October [3] E. N. Mazzae, F. Barickman, G. H. S. Baldwin, and G. Forkenbrock, Driver Crash Avoidance Behavior with ABS in an Intersection Incursion Scenario on Dry Versus Wet Pavement, in SAE Technical Paper, no , SAE International,

178 APPENDIX B KRUSKAL-WALLIS TEST BOX PLOTS FOR BRAKE APPLICATION TIME Figure B.1: Left incursion, brake application time, Kruskal-Wallis box plot. 158

179 Figure B.2: Right incursion, brake application time, Kruskal-Wallis box plot. Figure B.3: Stopping vehicle, brake application time, Kruskal-Wallis box plot. 159

180 Figure B.4: Stopped vehicle, brake application time, Kruskal-Wallis box plot. Figure B.5: Scenario comparison, brake application time, Kruskal-Wallis box plot. 160

181 APPENDIX C SIMULINK BLOCK DIAGRAMS AND MATLAB CODE USED FOR DBS TESTING The following section provides all Simulink diagrams and code, for embedded MATLAB functions, used to simulate the prototype DBS system presented in this document. The diagram presented, in Figure C.1, is the top level diagram of this simulation and contains several blocks that perform different functions in the simulation. The left most block generates a driver brake pedal force profile based on either time series data for a previously recorded driver response or frequency content and brake application time parameters for a stochastic frequency content based input. Working from left to right the next block contains the prototype DBS system, this block takes driver pedal force as an input and outputs the system command pressure which is either commanded by the driver or the DBS system. The next block to the right converts the pressure from the DBS system, which operates on PSI, to a pressure input for the TruckSim model which used MPa. The block directly below this is a throttle override which is used to ensure that the TruckSim speed controller does not apply the throttle during braking maneuvers. This data is then input into the TruckSim S-Function in order to run the simulation. The output from the TruckSim block is sent to the workspace with the variable named Simulation Results. Also, during the simulation the vehicle forward velocity is monitored to determine when the vehicle has stopped. The simulation is stopped once this state is achieved. 161

182 162 Figure C.1: Top level of Simulink diagram used for DBS simulation.