COMMERCIAL DRIVER BEHAVIORAL ANALYSIS USING REAL-TIME MIRROR CHECK MONITORING TOOLS

Transcription

1 COMMERCIAL DRIVER BEHAVIORAL ANALYSIS USING REAL-TIME MIRROR CHECK MONITORING TOOLS Margaret McNamara Purdue University 55 Stadium Mall Drive, West Lafayette, IN 4796 Tel: Zackary Knoll Maven Machines, Inc Liberty Ave, Pittsburgh, PA Tel: Clara Voss Purdue University 55 Stadium Mall Drive, West Lafayette, IN 4796 Tel: Darcy Bullock Corresponding Author Purdue University 55 Stadium Mall Drive, West Lafayette, IN 4796 Tel: /31/216

2 McNamara, Knoll, Voss, Bullock ABSTRACT Commercial truck drivers often operate vehicles for long periods of time and there has been significant effort to develop rules and best practices that minimize fatigue and ensure alert drivers. A new metric has emerged that is based in Federal Motor Carrier Safety Administration (FMCSA) guidelines for mirror check rate, measured in real-time on the road with an instrumented headset. While driver workload has traditionally been measured through eye glances in driving simulators and instrumented vehicles, the headset can be deployed to drivers for use in everyday driving situations with minimal obstruction. The drivers mirror check rate is recorded in real time along with a timestamp and GPS location, and these events can be used to identify individual driver behavior and opportunities for improvement. This paper performs an exploratory analysis of this new data set on a segment of I-7 in Ohio, West Virginia, and Pennsylvania. The dataset included 19 drivers operating 191 trips, spanning over 357 hours and 14, miles of driving. Data per individual driver was processed to develop mirror check performance measures for comparing rural and urban driving, check rate by hour of day, and cell phone use. The results show that mirror check rates are heavily affected by individual driver factors. The paper concludes by suggesting these mirror check performance measures can be an effective tool for companies to monitor driver alertness and to identify drivers that might benefit from additional training. Keywords: Commercial drivers, Driver performance, Mirror check rate, Cell phone 1

3 McNamara, Knoll, Voss, Bullock INTRODUCTION Commercial truck drivers often operate vehicles for long periods of time and there has been significant effort to develop rules and best practices that minimize fatigue and ensure alert drivers (1). Many studies have measured driver attention in various situations, mostly through the use of driving simulators and naturalistic studies (2). In general, those studies have relatively complex instrumentation that does not scale well to commercial carriers developing performance monitoring systems for their drivers. One such metric that shows promise for company administered performance monitoring is based upon a component of the Smith Safe Driving System (SSDS), which emphasizes checking the mirrors at regular intervals, which the Federal Motor Carrier Safety Administration (FMCSA) defines as at least every 5-8 seconds (3). In order to measure compliance with this mirror checking system, technologies are emerging that use a smart headset to track driver head motions and calculate the mirror check rate in real time. The headset pairs with the driver s phone and can report head turn events along with a timestamp and GPS data so that the results can be plotted on a map. Fleet managers can use these data to identify long-term behavioral characteristics of drivers and opportunities for improvement. These data can also be used in aggregate to examine the performance of multiple drivers along the same route. Unlike traditional driving simulators and instrumented vehicle tests, these data come from real-world driving wherein drivers may encounter situations that are difficult to replicate in a controlled experiment. LITERATURE REVIEW Driver workload and visual attention are prominent subjects in the literature. Studies such as Danno, Kutila, and Kortelainen use cameras to determine a driver s useful field of view (UFOV) both in simulation and field driving situations (4). Driving simulators are often used in these studies for their ability to provide a controlled experiment and test reaction to events that are rare on the road. Samuel and Fisher evaluated the minimum forward glance duration necessary to detect threats in front of the driver when conducting tasks inside the vehicle, recommending a minimum of 7 seconds (5). Driver workload that is too low or too high can be related to higher crash rates, and Brookhuis and de Waard used driver simulations and physiological measures such as heart rate and brain activity to monitor it (6). The dangers of low workload have been well-studied in the aviation industry, with extensive training and warnings in place to ensure that pilots remain alert when autopilot is engaged (7). Eye glance studies have found that mirror checks increase during the decision phase for a lane change maneuver (8). The literature is still emerging on the effect of cell phone use on driver distraction, with experimental research supporting that the addition of conversation increases the workload of drivers (9, 1). Naturalistic driver studies, however, have not found that cell phone use increases risk of a safety-critical event, and commercial motor vehicle drives are often at decreased risk while talking on a hands-free set (11). Studies of commercial vehicle drivers are quite heavily focused on the regulations surrounding the hours of service. Chen and Xie used machine learning on operating hours and crash data to determine optimal daily schedules for drivers to manage fatigue (12). Commercial drivers differ from the general driving population in both vehicle characteristics and driver experience, and applying results from general studies to commercial drivers may mischaracterize their qualities. Surrogate safety measures, such as mirror check rate, have not been specifically studied for commercial vehicle drivers in the literature. STUDY DATA AND OBJECTIVE This paper contains an initial exploration of a data set collected on I-7 between Columbus and the Pennsylvania-Maryland border that monitors head movements of drivers associated with mirror checks. This I-7 corridor was selected due to its high percentage of truck traffic and frequency of use by drivers wearing the headsets. The data set contained anonymized information collected over the six-month period. For each data point, the mirror check rate per minute was calculated using a two-minute rolling window. In order to relate the GPS data to roadway information and compare the drivers on the same route, the data points were linearly referenced along the I-7 corridor. The data were also segmented into separate trips using the time between messages, with gaps longer than an hour signaling the start of a new trip. The resulting dataset included 19 drivers operating 191 trips, spanning over 357 hours and 14, miles of 2

4 McNamara, Knoll, Voss, Bullock driving. The objectives of this exploratory study were to: 1. Characterize the variation in mirror checks among drivers and compare with the FMCSA recommendations. 2. Compare rural and urban areas of I-7 to determine if there is a difference in driver behavior between rural and urban areas on mirror check rates. 3. Examine mirror checks by time of day to determine if time of day influences mirror check rate. 4. Compare driver mirror check rates when talking on a cell phone against baseline. HEADSET TECHNOLOGY The data collection device is a hands-free Bluetooth headset used by commercial truck drivers. The device utilizes microelectromechanical systems (MEMS) technology to capture head motion in 6 degrees of freedom at a rate of 5 times per second. Accelerometer and gyroscope sensor data are transmitted to a host cell phone in the vehicle via Bluetooth. In addition to the 6 degrees of freedom inertial data, GPS on the phone is used to attach geographic location to each data point along with a flag indicating if a cell phone call was active. Data Processing and Characterization Methodology Once the data are sent to the phone, algorithms extract a set of discrete events which include left and right mirror checks. These events are tagged with GPS location and timestamp. The company has tested these algorithms by comparison with ground truth video data from the field. With a limited set of drivers and around 4 hours of video, the data has been verified to 98% accuracy. The 2% error is mostly comprised of false positives on the left side, resulting from the more subtle motion that is characterized by a left check. The data presented in this paper were gathered from January 1 - June 2, 216. It characterizes commercial truck driver performances and behaviors along the I-7 corridor between Columbus, Ohio, and the Pennsylvania-Maryland border, with most of the activity taking place between Columbus and the junction with I-79 south of Pittsburgh, Pennsylvania. The data set is comprised of a time ordered series of events generated by each specific driver. Each event is tagged with a UTC timestamp, a unique driver ID, and a latitude and longitude. Anonymous driver IDs were assigned in accordance with the Purdue University Human Research Protection Program (HRPP) approved protocol. Each event may also include a unique set of parameters to characterize details about the event, accompanied by a status. The status contains critical information regarding whether the vehicle is in motion and whether or not the device was being properly worn. This information was used to ensure that only valid driving data was used in the analysis presented in this paper. INDIVIDUAL DRIVER MIRROR CHECK RATES Figure 1 depicts an eastbound trip by Driver 22 through Pennsylvania on I-7 and the Pennsylvania Turnpike (I-7/76) on May 2, 216. Figure 1b charts the driver s mirror check rate per minute along the route, referenced in miles from the Indiana-Ohio border. The FMCSA guideline mirror check interval of 5-8 seconds, or checks per minute, is shaded in yellow. The general trend of mirror check rate is decreasing, with the peak mirror check rate occurring near the interchange between I-7 and the Pennsylvania Turnpike at mile 279 in Figure 1b. Another way to characterize the mirror check data is to examine the gaps between individual mirror checks. Over this trip, 31% of the mirror checks were within the FMCSA guideline of 5 to 8 seconds and 17% were after a gap of more than 15 seconds. 3

5 Mirror Checks per Minute McNamara, Knoll, Voss, Bullock a) Subsegment of route b) Mirror check rate along route, Driver 22 Figure 1 Example trip with mirror check rate Mile Along Route Each driver s mirror check patterns can be aggregated over many trips and summarized, as in Figure 2. This Pareto-sorted graph shows the length of the gap between mirror checks for each driver, separated into four categories: greater than 15 seconds, 8 to 15 seconds, 5 to 8 seconds (the recommended threshold), and less than 5 seconds. The last category includes very quick left and right mirror scans that could otherwise be counted as a single check event, though there is no algorithm in place at the moment to connect them. From the figure, Driver 1 appears to have the lowest compliance, with nearly 6% of mirror checks happening at a gap greater than 8 seconds. Driver 14, on the other hand, only has 12% of checks after gaps longer than 8 seconds. 4

6 McNamara, Knoll, Voss, Bullock % 9% 8% 7% 6% 5% 4% 3% 2% 1% % > 15 s 8-15 s 5-8 s < 5 s Driver Figure 2 Time between mirror checks, all drivers, Pareto sorted Figure 3a shows the check distribution for Driver 8, with the same coloring as Figure 2, aggregated from 2 trips and over 1,65 miles driven. This driver can be compared to Drivers 2 and 18 in Figure 3b and c, respectively. The longer tails of the latter two drivers indicate a higher proportion of checks that take place more than 8 seconds apart. Driver 8 also has more gaps less than 5 seconds than the other two drivers, indicating a tendency to check both mirrors in a quick scan. Figure 2 shows that Driver 1 had the highest proportion of check gaps longer than 15 seconds. As shown in Figure 3e, over the 1 mile trip, this driver has a more spread out distribution than Driver 8, with more outliers on the tail of the distribution. This suggests that Driver 1 may be a good candidate for additional training, though it may also be an artifact of the limited sample size of this data set. The simple graphical representation provides a mechanism to identify drivers who may benefit from intervention as well as those who may be used as examples of best practice. 5

7 Frequency Frequency Frequency Frequency McNamara, Knoll, Voss, Bullock Time between mirror checks [s] a) Time between mirror checks, Driver 8 (1,65 miles) Time between mirror checks [s] b) Time between mirror checks, Driver 2 (62 miles) Time between mirror checks [s] c) Time between mirror checks, Driver 18 (3,7 miles) Time between mirror checks [s] d) Time between mirror checks, Driver 1 (1 miles) Figure 3 Time between mirror checks analysis 6

8 McNamara, Knoll, Voss, Bullock CORRIDOR ANALYSIS OF DRIVER MIRROR CHECK RATES The trips by all drivers in the dataset can be aggregated to evaluate their spatial characteristics along the roadway. Figure 4a depicts the route studied, from the west side of Columbus, Ohio, to where I-7 meets the Pennsylvania-Maryland border. The beginning and ending reference miles are also shown, with beginning at the Indiana-Ohio border for ease of measurement. Figure 4b and c shows the medians from all trips in the eastbound and westbound directions, with state boundaries and the FMCSA guidelines indicated. The median check rate is largely within the checks per minute guideline. There is some anomalous peaking around the Ohio-West Virginia border, likely due to many drivers taking I-47 around Wheeling, West Virginia. The check rate peaks around urban areas such as Columbus Ohio, from mile 92 to 18, but there is not a strong differentiation in mirror checks between urban and rural segments of the route. 7

9 Mirror Checks per Minute Mirror Checks per Minute McNamara, Knoll, Voss, Bullock a) Map of entire I-7 route OH WV PA Mile Along Route a) Eastbound OH WV PA Mile Along Route b) Westbound Figure 4 Median mirror check rate along route, all drivers Figure 5a shows a subset of the route with the mile measurements around Columbus and Zanesville, Ohio marked. Figure 5b shows the number of eastbound trips on which Driver 18 drove each mile over the course of the study. The lower sample size around Columbus suggests that the driver often takes the bypass route, and anomalies like that at callout i could be due to a rest stop in a segment that lasted more than one hour and was counted as two trips. Figure 5c shows a box-and-whisker plot of Driver 18 s check rate along this segment of the route. Callout i denotes the driver passing through Columbus, where the mirror check rate is fairly consistent and within the FMCSA guideline. Callout ii shows an increase in the interquartile range of mirror check rate as the driver enters a rural stretch of road. Callout iii shows another case of very wide variation in mirror check rate as the driver approaches Zanesville, Ohio. This could be attributed to variation in fatigue levels after a 8

10 McNamara, Knoll, Voss, Bullock forty-mile stretch of rural driving. Finally, callout iv shows that the interquartile range through the interchanges at Zanesville is not as tight as that around Columbus. To examine another possible cause for this wide variation in mirror check rate, a speed profile is presented in Figure 5d. This graph is derived from crowdsourced probe vehicle speeds on the road, representing the conditions of the surrounding traffic at the times when the driver passed each mile marker (13). The slowest traffic conditions correspond to the interchanges between I-7 and I-27 around Columbus. The rural stretch from mile 11 to 16 was largely experienced at free-flow speeds, aside from a few minutes at mile 142 that were below 15 mph. This anomaly could partially explain the variation in check rate at that mile marker, but from this sample there appears to be little correlation between the speeds on the road and the driver s mirror check rate. 9

11 Minutes Mirror Checks er Minute Trips McNamara, Knoll, Voss, Bullock a) Subsegment of route near Columbus, OH i b) Number of trips by mile along route, Driver 18 Miles Along Route Columbus i ii iii iv Zanesville Miles Along Route c) Box-and-whisker plot of mirror check rate for Driver mph mph mph mph mph mph > 65 mph d) Speed profile for route subsegment Mile Along Route Figure 5 Mirror check rate along route subsegment for one driver 1

12 McNamara, Knoll, Voss, Bullock TIME OF DAY ANALYSIS OF DRIVER MIRROR CHECK RATES The time of day is another factor that can influence driving behavior through fatigue, traffic, and varying light levels. Figure 6a shows the mirror check rate by hour of day for all drivers, in local time. While the range in each hour is very large and no outlier analysis was done, the interquartile range of most hours is very tight. From 7 to 18, the central boxes are near the guideline values of checks per minute. From 19 to, the interquartile ranges increase and mirror check rates drop with the sun setting and late-day fatigue. To further examine individual driver characteristics, Figure 6b-d show the same graph for Drivers 8, 2, and 18. Driver 8 works very consistently between 7 and 2 with the median check rate above 12 checks per minute for every hour but the last at 2. Driver 2 has a much different curve, partially due to a smaller but still significant sample size. The 8 hour is likely spent driving to the pickup point, and the median value is below the guideline. The rest of the day shows highly variable mirror check rates, particularly around 19. Driver 18 s trips spanned a longer portion of the day than the other two, and even in the early part of the day the interquartile range is small, though lower than Driver 8 s. The two large jumps in interquartile range at 8 and 19 may coincide with rush hour and greater amounts of merging and weaving traffic. 11

13 Mirror Check Rate Mirror Check Rate Mirror Check Rate Mirror Check Rate McNamara, Knoll, Voss, Bullock Hour of Day a) Mirror check rate by hour of day, all drivers Hour of Day b) Mirror check rate by hour of day, Driver Hour of Day c) Mirror check rate by hour of day, Driver Hour of Day d) Mirror check rate by hour of day, Driver 18 Figure 6 Mirror check rate by hour of day 12

14 McNamara, Knoll, Voss, Bullock ANALYSIS OF CELL PHONE USAGE AND DRIVER MIRROR CHECK RATES To further examine the effect of driver behavior on workload and attention, the mirror check rate may be plotted by cell phone call status, as in Figure 7. These graphs are cumulative frequency diagrams, with higher numbers denoting more frequent checks and the FMCSA guideline is highlighted. A curve that is farther to the right indicates higher mirror check rates. As discussed earlier, the state of the literature is divided on the effects of cell phone use, and this exploration of the data set is equally divided. Driver 8 has consistently high performance whether on the phone or not, as seen in Figure 7a. Driver 2 (Figure 7b) checks his mirrors more often when on the phone than not, with a median value of 8 checks per minute meeting FMCSA guidelines. On the other hand, there is a significant left shift in the phone use curve for Driver 18 in Figure 7c, demonstrating that phone use significantly decreases his mirror check rate. The relative sample size for these cases are also of note. Drivers 8 and 2 spent 69% and 76% of sampled driving time, respectively, on the phone. Driver 18 spent only 14% of the time driving on the phone, possibly indicative that the driver knows the impact of cell phone use on his driving ability and self-regulates it when possible. In contrast, Drivers 8 and 2 may be comfortable speaking on the phone, with Driver 2 even seeing some benefit in alertness from it. 13

15 Cumulative Frequency Cumulative Frequency Cumulative Frequency McNamara, Knoll, Voss, Bullock On phone Not on Phone Mirror Check Rate a) Driver 8 On phone Not on Phone Mirror Check Rate b) Driver 2 On phone Not on Phone Mirror Check Rate c) Driver 18 Figure 7 Cumulative frequency diagrams of mirror check rate by phone use CONCLUSIONS AND FUTURE RESEARCH Driver inattention is a critical factor in most crashes. However, identifying cost effective mechanisms for assessing the driver s attention level in commercial vehicles has been a challenge. One such metric that shows promise for company administered performance monitoring is based upon a component of the Smith Safe Driving System (SSDS) and FMCSA regulations that emphasize checking the mirrors at least every 14

16 McNamara, Knoll, Voss, Bullock seconds. This paper presents a methodology that uses a smart headset designed to track driver head motions and calculates the mirror check rate in real time. The headset pairs with the driver s phone and can report head turn events along with a timestamp and GPS. The three study objectives addressed individual compliance rates as well as roadway, time of day, and cell phone use effects. The conclusions were: 1. From a driver population of 19 drivers, that drove over 14, miles on I-7 over a six month period, it was found that driver conformance with the SSDS ranged from 42% to 88% (Figure 3). 2. With regards to influence of rural and urban driving on I-7, no significant differences were observed in the overall driver population. 3. With regards to time of day, some modest reduction in mirror checks were observed by some drivers late in the day, though no trend was visible in the overall dataset. 4. With regards to cell phone use, drivers experience different effects on mirror check rate and may self-regulate their phone habits because of it. This preliminary analysis suggests that additional human subject protocols need to be developed to expand this study to include additional variables, particularly relating to individual driver demographics. Applying this analysis over a longer term period will also allow more sophisticated econometric analysis of factors influencing mirror check rate to be developed. There are many other factors that can influence the mirror check rate that were not accounted for in this study, such as the presence of work zones or weather. Future analysis can be done to include these variables that can further explain the fluctuations in mirror check rate. ACKNOWLEDGEMENTS This work was supported in part by the Purdue University Andrews Fellowship, INRIX, Maven Machines, Inc. and the Joint Transportation Research Program and the Indiana Department of Transportation (INDOT). The probe vehicle data used in this study was provided by INRIX. The headset data was provided by Maven Machines, Inc. The contents of this paper reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein, and do not necessarily reflect the official views or policies of the sponsoring organizations. These contents do not constitute a standard, specification, or regulation. REFERENCES 1. FMCSA. Regulations. Federal Motor Carrier Safety Administration, March 31, Accessed July 26, Ranney, T. A. Driver Distraction: A Review of the Current State-of-Knowledge. National Highway Traffic Safety Administration, Washington, D.C., Final Report DOT HS , FMCSA. CMV Driving Tips - Inadequate Surveillance. Federal Motor Carrier Safey Administration, March 31, cmv-driving-tips-inadequate-surveillance. Accessed July 1, Danno, M., M. Kutila, and J. M. Kortelainen. Measurement of Driver's Visual Attention Capabilities Using Real-Time UFOV Method. International Journal of Intelligent Transportation Systems, Vol. 9, no. 3, 211, pp DOI: 1.17/s Samuel, S., and D. L. Fisher. Evaluation of the Minimum Forward Roadway Glance Duration. Transportation Research Record, no. 2518, Transportation Research Board, Washington, D.C., 215, pp Brookhuis, K. A., and D. de Waard. Monitoring drivers' mental workload in driving simulators using physiological measures. Accident Analysis and Prevention, Vol. 42, 21, pp DOI: 1.116/j.aap Michaels, D., and A. Pasztor. Aviation Experts Urge Caution Releasing Self-Driving Cars. The Wall 15

17 McNamara, Knoll, Voss, Bullock Street Journal, July 27, aviation-experts-suggest-caution-releasing-self-driving-cars Accessed July 31, Tijerina, L., W. R. Garrott, D. Stoltzfus, and E. Parmer. Eye Glance Behavior of Van and Passenger Car Drivers During Lane Change Decision Phase. Transportation Research Record, no. 1937, Transportation Research Board, Washington, D.C., 25, pp Tokunaga, R. A., T. Hagiwara, S. Kagaya, and Y. Onodera. Cellular Telephone Conversation While Driving: Effects on Driver Reaction Time and Subjective Mental Workload. Transportation Research Record, no. 1724, Transportation Research Board, Washington, D.C., 2, pp Laberge, J., C. Scialfa, C. White, and J. Caird. Effects of Passenger and Cellular Phone Conversations on Driver Distraction. Transportation Research Record, no. 1899, Transportation Research Board, Washington, D.C., 24, pp Fitch, G. M., K. Grove, R. J. Hanowski, and M. A. Perez. Compensatory Behavior of Drivers When Conversing on a Cell Phone: Investigation with Naturalistic Driving Data. Transportation Research Record, no. 2434, Transportation Research Board, Washington, D.C., 214, pp Chen, C., and Y. Xie. Machine Learning for Commercial Large Truck Drivers' Driving Pattern Recognition. Transportation Research Record, no. 2517, Transporatation Research Board, Washington, D.C., 215, pp McNamara, M., H. Li, S. Remias, L. Richardson, E. Cox, D. Horton, and D. M. Bullock. Using Real-Time Probe Vehicle Data to Manage Unplanned Detour Routes. ITE Journal, Vol. 85, no. 12, December 215, pp