An Evaluation Study of Driver Profiling Fuzzy Algorithms using Smartphones

Transcription

1 An Evaluation Study of Profiling Fuzzy Algorithms using Smartphones German Castignani University of Luxembourg / SnT german.castignani@uni.lu Raphaël Frank University of Luxembourg / SnT raphael.frank@uni.lu Thomas Engel University of Luxembourg / SnT thomas.engel@uni.lu Abstract Profiling driving behavior has become a relevant aspect in fleet management, automotive insurance and eco-driving. Detecting inefficient or aggressive drivers can help reducing fleet degradation, insurance policy cost and fuel consumption. In this paper, we present a Fuzzy-Logic based driver scoring mechanism that uses smartphone sensing data, including accelerometers and GPS. In order to evaluate the proposed mechanism, we have collected traces from a testbed consisting in 2 vehicles equipped with an Android sensing application we have developed to this end. The results show that the proposed sensing variables using smartphones can be merged to provide each driver with a single score. I. INTRODUCTION Driving behavior profiling has become relevant in different environments. For instance, in the fleet management domain fleet administrators are interested in fine-grained information about the fleet usage, which is conditioned to the different driver usage patterns. On the car insurance market, Usage- Based Insurance (UBI) or Pay-As-You-Drive (PAYD) schemes aim to adapt the insurance policy cost to the individual driver behavior. In order to track the driver behavior, existing systems have been based on dedicated telematics boxes (e.g., Ingenie [6], Fairpay [12]) that log different parameters and driving events. The information logged by these boxes can be then manually retrieved or sent over the Internet through a wireless connection. However, the main drawbacks of these boxes are the high installation, operation and maintenance costs, requiring a high investment and limiting their wide deployment. Moreover, due to the increasing sensing capacities and the proliferation of mobile devices like tablets and smartphones (e.g., accelerometers, magnetometers, GPS), in the last years some application-based driver scoring tools have been proposed. In the car-insurance market, Aviva RateMyDrive [2] and StateFarm Feedback [11] appear as the most popular mobile applications for ios and Android. In both cases, the provided score is used as an input for adjusting the insurance policy cost, providing up to 2 % policy reduction. On the other hand, Greenroad [4] is an online platform for fleet management. In this platform, drivers use the sensing application and regularly send driving traces to the system which aggregates metrics from different drivers to provide fleet administrators with a description of individual and fleet /13/$31. c 213 IEEE riskiness and eco-driving related information (e.g., fuel consumption, CO 2 emissions). However, driving profiling through smartphones poses some issues about how different sensing parameters are considered to come out with a reliable score. In this paper, we review existing driver scoring techniques based on smartphone sensing data. Due to the imprecision of sensing data, we propose a fuzzy-logic mechanism to score drivers based on overspeed, acceleration and steering events. Moreover, we apply the proposed scoring algorithm to a driver sensing dataset we have obtained through an Android sensing tool we have developed to this end. This data-set consists in more than 55 trips from 2 different drivers of a Luxembourg logistics company. The paper is organized as follows. In Section II we introduce the related work. Then in Section III we model the proposed scoring mechanism using a Fuzzy Inference System. In Section IV we present the results for a large data-set we have collected using our own Android sensing application. Finally in Section V, we conclude the paper. II. RELATED WORK Driving profiling systems based on smartphone sensing data have been proposed so far. Eren et al. [3] designed a driver classification algorithm that distinguish between risky and safe drivers. They considered smoothed acceleration, gyroscope and magnetometer data from smartphones to detect start and end time of driving events (e.g., sudden maneuvers, aggressive steering, braking or acceleration) using a moving average algorithm and empirical thresholds. The authors computed the similarity of each event with template data (i.e., for risky and safe behavior) using Dynamic Time Warping (DTW) and used Bayesian classification to decide whether the driver is risky or safe. An evaluation study is proposed for fifteen drivers using iphones and fixed departure and arrival points showing a successful classification rate of 93.3%. Johnson et al. [7] also proposed a DTW-based driver profile algorithm using smartphone sensors, GPS and camera called MIROAD. They evaluated in this work the performance of different sensor fusion sets to detect lateral and longitudinal movements. After an evaluation over 2 driving events, the authors showed that the sensor fusion set composed of the x-axis rotation rate, y-axis acceleration and pitch provide the best classification performance using DTW.

2 Paefgen et al. [9] focused on the precision of smartphone sensing data for the analysis of driver behavior mainly oriented to the insurance market. After a calibration process where the user manually sets the principal direction of the vehicle, the mobile application starts collecting acceleration, braking and steering events. These events are triggered if the sensing data surpass some predefined thresholds (i.e.,.1g for acceleration and braking and.2g for steering). The authors proposed a measurement study to compare event detection using smartphone sensors against a fixed telematic box. They observed that the event count distribution obtained corresponded to different statistical distributions, which was mainly due to diverse smartphone-to-car fixing and shifting. However, the authors found some correlations between smartphone and fixed box events and described some possible sources of error. You et al. [13] proposed CarSafe, a smartphone application that fuses information from front and rear cameras, sensors and GPS to detect dangerous driving events. In particular, the authors showed that drowsiness (the main cause of car accidents [8]) can be detected using the front camera and image processing algorithms with an accuracy of 85 %. With the aim of providing drivers with useful hints to reduce energy consumption, Araujo et al [1] have proposed a smartphone application that combines GPS and CAN-bus information (using an OBD device). Some of the possible hints are to switch-off the engine, to shift gears earlier or to decelerate. As input data, they consider average, minimum and maximum values for speed, acceleration and fuel consumption and they combine them in a Fuzzy-System. They evaluated and validated their algorithms in an Android platform used in several experiments in a single car. III. FUZZY INFERENCE SYSTEM FOR DRIVER PROFILING As introduced before, existing driver profiling mechanisms are based on multiple input data and classification mechanisms. Commercial applications as Greenroad [4] rely on GPS and smartphone sensor data to find events. In this application, the score is then simply calculated as an event rate, i.e., the number of events per unit of distance that the application has counted. In this case, all type of events have the same priority for the scoring and are merged in a global event counter. In this paper, we aim to present a new platform only based on smartphone data that uses a Fuzzy-System combining input data and multiple inference rules and come out with a reliable score. The architecture of the system is briefly illustrated in Figure 1. Sensing data from the smartphone is filtered and events are detected for the different metrics. Then, input data is fuzzified and fuzzy rules are applied in a Fuzzy Inference Engine. Finally, the defuzzification process allows us to provide a score (from to 1). In this paper, the proposed fuzzy mechanism has been evaluated in a large measurement study counting more than 55 trips of a logistic fleet. A. Sensor Data and Input Variables We consider different sensor data as input for the proposed classification system. This data is based on input from four Fig. 1: Fuzzy Inference System sensors: GPS, accelerometer, magnetometer and gravity sensor. The different input variables are described below. 1) Total Acceleration: We define the total acceleration as the magnitude of the acceleration vector measured from the smartphone accelerometer (a D in m/s 2 ) in the device coordinate system after removing the gravity component. The X component of this vector points to the right side of the smartphone (in portrait position), the Y components points to the top of the phone and finally the Z component is perpendicular to the screen of the device. In order to transform the acceleration vector to the earth s coordinate system, a E, the application calculates the rotation matrix (R) and computes the product a E = R a D. The total acceleration is computed as the norm of the acceleration vector (ka E k). Note that if the total acceleration is used as a metric, it is not necessary to convert the acceleration vector to a different coordinate system. As input for the classifier, we have considered two variables: SA M and SA A, i.e., the number of moderate and aggressive acceleration events per kilometer respectively. A moderate acceleration event is triggered for kak > 1.5m/s 2 while an aggressive acceleration event occurs when kak > 3m/s 2. These values have been obtained from previous empirical studies [8], [9] and the analysis of our validation experiments (see Section III-B). Note that we have considered sensor acceleration as a total vector magnitude since, due to noises in the accelerometer signal (e.g., orientation change, phone manipulation), it is difficult to distinguish longitudinal and lateral accelerations in a reliable manner. 2) Linear Acceleration: In order to obtain a magnitude of the acceleration and braking in the vehicle s displacement axis we consider the GPS speed variation. Differently from smartphone sensors, which have a high sampling rate (between 1 Hz and 1 Hz), linear acceleration from GPS is provided with a sampling rate equal to 1 Hz. For the classifier, we have considered four input variables based on the GPS linear acceleration, i.e., the speed variation

3 rate (in m/s 2 ) between two GPS samples. The first two variables are the maximum positive (GA P ) and negative (GA N ) linear acceleration. Second, we consider the number of moderate (GA M ) and aggressive (GA A ) events per kilometer. Moderate linear acceleration events are those which absolute value is greater than 1m/s 2 and aggressive events are those with values greater than 2.5m/s 2. 3) Overspeed: Overspeed is a useful metric to characterize the driver behavior. For the best of our knowledge, no prior work on driver profiling has considered overspeed as input variable. In order to compute overspeed, we calculated the difference between the vehicle s speed at each location update and the speed limit for that specific location. As for the speed limit, we can obtain the speed limit per location by accessing a web-service based on OpenStreetMap 1. We considered three input variables for overspeed. First, the relative overspeed time OS T is the normalized amount of time the driver incurs in overspeed, i.e., OS T =indicates no overspeed and OS T =1indicates that the driver has incurred in overspeed during the complete duration of the trip. We also consider the average overspeed OS A and the maximum overspeed OS P per driver. 4) Steering rate: In order to have a measure of the steering behavior of drivers, we consider the bearing angle provided by the GPS. The bearing indicates the direction of the vehicle as the relative angle (from to 36 ) to the north. Note that steering rate may also be quantified using internal gyroscope and magnetometer but, in our experiments, we have observed a high level of noise in this sensing data. Then, we computed the steering rate as the variation of the bearing angle for two consecutive GPS location updates (in /s). In the case of steering, we consider three input variables for the fuzzy system. First, the number of moderate and aggressive events per kilometer, BE M and BE A respectively. Based on our empirical findings, moderate steering events (BE M ) are those where the bearing rate is greater than 1 /s and aggressive events (BE A ) implies a bearing rate greater than 4 /s. Moreover, we consider as an input the maximum observed steering rate (BR P ) for each driver. B. Validation of sensing variables In order to validate the smartphone-based sensor data we have set up a testbed using real time information from a vehicle using a PCAN USB adapter [1], which provides CAN-bus data at a very high sampling rate and high accuracy. The experiment consisted in collecting traces from a Smart ED electric vehicle during several trips in a suburban area of Luxembourg. Traces have been collected in parallel using a Samsung Gio smartphone (with our sensor tool) mounted in a car-holder and a Dell XPS laptop directly connected to the PCAN USB adapter. For the traces collection, both devices have been synchronized in time and traces have been analyzed offline. In Figure 2 we present the main results of this experimentation. Even if the PCAN adapter does not provide CALM AVERAGE MODERATE AGGRESSIVE Fig. 3: Output Fuzzy Set acceleration information, we obtained the throttle and brake pedals position over the time (see Figure 2a). On the other hand, we obtained for the same trip the total acceleration from the internal sensors and the GPS acceleration and braking using the smartphone. Note that the sensor data in Figure 2b corresponds to raw/unfiltered data. Then, to be considered as input for the classification system, this data is filtered using a Kalman filter. We observe that high acceleration and braking events in Figure 2a (obtained with the PCAN adapter) are correlated with sensor and GPS data. Note for example the hard braking events between time 8 s and 13 s that are observed both in Figure 2b and 2c. Also between time 32 s and the end of the trip, a sequence of hard acceleration and braking events is observed in all the figures. Regarding the steering traces, we were able to log the steering wheel position of the car (see Figure 2d) with a high sampling rate. We observe that there is a high correlation of the steering wheel position and the GPS bearing rate calculated using the smartphone s GPS. Observe for example in both figures the steering events at time 5 s, 13 s and 22 s, where the steering wheel position drastically changes and it is reflected in the GPS bearing variation in Figure 2e. C. Fuzzy Sets In the proposed scoring mechanism, a driver is characterized by a set of variables describing overspeed (OS T,OS A,OS P ), GPS acceleration and braking (GA P,GA N,GA A,GA M ), sensor acceleration (SA M,SA A ) and steering (BE A,BE M,BE P ). These input variables are assigned with a non-numeric linguistic value to facilitate rules definition. Also, we define an output variable to represent the driver score (S i ). The fuzzy sets for the input variables are triangular and in all cases consider three linguistic values: Low, Medium and High (L, M, H). The limits of each sets have been carefully defined based on empirical data and previous studies [8]. Regarding the output variable, we consider four categories for the score: Calm, Average, Moderate and Aggressive (CA, AV, MO, AG). The fuzzy set for the output variable is illustrated in Figure 3. D. Fuzzy Rules and Scoring We have defined a set of Fuzzy rules that combine the different input variables in order to provide a score. First, we define rules for all possible combinations (i.e., up to 27) of

4 Throttle pressure (%) Brake pressure (%) Total XY True Acceleration (m/s2) Acceleration (m/s2) (a) PCAN throttle and brake pressure (b) Sensor total acceleration (c) GPS acceleration Steering Angle Bearing Rate (degrees/s) (d) PCAN steering wheel position (e) GPS steering rate Fig. 2: Comparison between OBD and smartphone sensing variables overspeed data (OS T,OS A,OS P ). For example: OS T = L & OS A = L & OS P = L ) CA indicates that drivers incurring in a low overspeed time and a low average and maximum overspeed are Calm drivers. Second, moderate and aggressive sensor acceleration events are combined in 9 rules (e.g., SA M = M & SA A = M ) MO). Also, GPS acceleration and braking data are combined in two subsets of 9 rules each. In the first subset moderate and aggressive GPS acceleration events are combined. In the second subset we consider all possible combinations of the maximum acceleration and braking magnitudes (measured in m/s 2 ). For instance, one of the rules indicates GA N = M & GA P = M ) MO, which indicates that a driver incurring in medium acceleration and braking magnitudes are classified as moderate drivers. As for the acceleration and breaking events, all the combinations of steering moderate and aggressive events are considered in 9 rules. Also, three rules are considered for the maximum steering rate, e.g., BE P = M ) AV. After fuzzifying the input data and evaluating the inference rules, the Fuzzy Inference Engine outputs a score based through a defuzzification process. For each evaluated rule, an output will be triggered with an associated membership degree. Then, a given driver will belong to one or more categories (CA, AV, MO, AG) with a membership degree. In the case that multiple rules outcome the same category, the minimum membership degree among these rules is chosen. An output signal is calculated by computing the logical sum of each category output. An illustration of this process is shown in Figure 4, in which two outputs are triggered, i.e., AV with membership.5 and MO with membership.2. The two outputs are then summed to obtain a final output signal and finally, in the defuzzification process, we consider the COG algorithm, which calculates the center of gravity of the area under the curve and outputs the score (S) as the coordinate x of this point. In the proposed scoring mechanism, we consider that a driver i has a score S i which is a combination of the score obtained when driving in urban/city areas (S c ), in suburban areas (S s ) and extra-urban areas (S e ). Finally, the driver score is calculated as a weighted sum of each individual score, i.e., S i =.5 S c +.3 S s +.2 S e. These weighs have been selected since aggressive driver behaviors are more risky in urban environments than in suburban and extra-urban environments, due for example to the presence of pedestrians in urban environments [5]. A. Android tool IV. EVALUATION STUDY The evaluation of the scoring mechanism consisted in the collection of the traces and the analysis of data and score

5 Fig. 4: Defuzzification and Scoring In the proposed testbed, each driver carried a Samsung Galaxy Gio S566 (running Android 2.3.2) smartphone in a holder mounted on the vehicle s windshield and collected traces that were regularly and automatically sent to a remote server in our lab through a mobile network connection. The traces were organized in our server by individual trips per driver and sensing data has been analyzed and filtered in order to obtain, for each driver, the input variables of the Fuzzy System. Fig. 5: Collected Trips Heatmap computation. In order to collect traces, an Android driver sensing tool was developed. This application consists in a background service and a simple user interface gathering accelerometer, magnetometer, gravity sensor and GPS receiver traces. The traces are gathered in an internal SQLite database that organizes the sensing data for each individual trip (the start and stop of each event is manually triggered by the driver). When a new trip starts, the application first calibrates the accelerometer (by extracting the offset) and gets a first GPS fix. When the application finishes the calibration process, the driver can start moving and the application logs the traces in the database. Regularly, the application can upload the collected trips to a central database through a WiFi or 3G network connection. B. Testbed In order to evaluate the scoring mechanism, we have deployed a testbed that involves 2 drivers using small commercial vehicles (VW Caddy and Renault Kangoo) of a Luxembourg logistics company. These vehicles have moved for 869 km counting 566 individual trips in urban, suburban and extra-urban areas of Luxembourg during five weeks. A geographical visualization of this traces is illustrated in Figure 5, showing a heatmap that indicates the number of traces collected in each area of the country (i.e., black and yellow indicates minimum and maximum number of trips respectively). C. Results Figure 6 shows a subset of the input variables for the Fuzzy System that have been obtained for the different drivers in urban environments. Each input variable has been calculated by considering the total number of urban trips for each driver. We observe that the different input variables widely vary for the different drivers. Regarding overspeed metrics, drivers spent between 3.7% and 13.5% of the time incurring in overspeed in urban areas. These average speeds vary between 3.6 m/s and 8.5 m/s. In terms of steering, drivers incur in between 1.3 and 1.9 average events per kilometer, which can reach up to a maximum steering rate 279 /s. In the case of linear acceleration events measured with the GPS, we have observed up to 31 moderate and aggressive events per kilometer for two different drivers. We have computed the score for each driver using the proposed Fuzzy Logic mechanism. The obtained scores, that are calculated as a weighted sum of urban, suburban and extraurban scores (see Section III-D) are illustrated in Figure 7, showing that drivers scores are distributed between and 77.65, which mainly correspond to moderate and aggressive drivers in Figure 3. Regarding the scores obtained for the different driving environments, we have observed that if all the drivers are considered, the average score for urban areas is greater than for suburban and extra-urban areas, i.e., for urban, for suburban and for extra-urban areas. This is mainly due to the higher overspeed and acceleration events when driving in the city compared to the roadway and highway scenarios, where people drive smoothly (i.e., there are less acceleration/braking events than in urban environments) and tend to respect speed limits due to enforcement. V. CONCLUSION In this paper, we have presented a driver sensing and scoring mechanism using smartphones. In particular, we have developed a sensing tool that gathers accelerometer, magnetometer,

6 Relative Time Speed (m/s) Average Maximum Events / km Moderate Aggressive (a) OS T (b) OS A and OS P (c) BE M and BE A 3 Moderate Aggressive Deceleration Acceleration Steering Rate ( /s) Events / km Acceleration (m/s2) (d) BE P (e) GA M and GA A (f) GA P and GA N Fig. 6: Evaluation Testbed Sensing Data Score Fig. 7: Score gravity sensor and GPS data. We have also proposed a scoring mechanism based on fuzzy variables and rules that combines sensing data in order to give a single score to characterize the different drivers. A large evaluation study allowed us to obtain sensing data from different drivers in different trips. The calculation of the score over these traces showed that drivers mainly belong to moderate and aggressive categories. For the future work, we aim to validate the accuracy of the obtained scores in a second sensing campaign involving the same drivers that have been informed about their score from the first experimental campaign. In this second experiment, we will focus on the analysis of driving behavior changes, since all drivers will tend to drive more efficiently in order to reduce their scores. ACKNOWLEDGMENT The authors would like to thank EPT for the financial support and NETCORE for their contribution to the testbed. REFERENCES [1] R. Araujo, A. Igreja, R. de Castro, and R. Araujo. Driving coach: A smartphone application to evaluate driving efficient patterns. In 212 IEEE Intelligent Vehicles Symposium (IV), pages 15 11, June 212. [2] Aviva PLC. Aviva Drive [3] H. Eren, S. Makinist, E. Akin, and A. Yilmaz. Estimating driving behavior by a smartphone. In 212 IEEE Intelligent Vehicles Symposium (IV), pages , June 212. [4] Greenroad. Introducing the revolutionary GreenRoad Smartphone Edition [5] K. E. Heck and K. C. Nathaniel. Driving among urban, suburban and rural youth in california. Youth Development, page 11. [6] Ingenie Services Limited. Car insurance for young drivers [7] D. A. Johnson and M. M. Trivedi. Driving style recognition using a smartphone as a sensor platform. In Intelligent Transportation Systems (ITSC), th International IEEE Conference on, page , 211. [8] U. D. o. T. National Highway Traffic Safety Administration. The 1- car naturalistic driving study - phase II results of the 1-car field experiment. Technical Report DOT HS , 26. [9] J. Paefgen, F. Kehr, Y. Zhai, and F. Michahelles. Driving behavior analysis with smartphones: Insights from a controlled field study. page 1. ACM Press, 212. [1] PEAK-System Technik GmbH. PCAN-USB Pro CAN/LIN Interface for High-speed USB Pro.2..html, 213. [11] State Farm Mutual Automobile Insurance Company. State Farm Feedback [12] Towergate Underwriting Group Limited. Fair Pay a radical new approach to motor insurance [13] C.-W. You, M. Montes-de Oca, T. J. Bao, N. D. Lane, H. Lu, G. Cardone, L. Torresani, and A. T. Campbell. CarSafe: a driver safety app that detects dangerous driving behavior using dual-cameras on smartphones. In Proceedings of the 212 ACM Conference on Ubiquitous Computing, UbiComp 12, page , New York, NY, USA, 212. ACM.