Traffic Parameter Methods for Surrogate Safety Comparative Study of Three Non-Intrusive Sensor Technologies CARSP 2015 Collision Prediction and Prevention Approaches Joshua Stipancic
2/32 Acknowledgements Co-authors Luis Miranda-Moreno, Associate Professor, McGill University Nicolas Saunier, Associate Professor, Polytechnique Montréal Funding Provided in part by the Natural Science and Engineering Research Council
3/32 Outline Introduction Literature Review Methodology Site Selection Technology Selection and Instrumentation Data Analysis Results Conflict Analysis Temporal Speed Variation Lateral Speed Variation Conclusions
4/32 Introduction Traditional Methods In practice, achieving safe road networks is a common requirement Perceptions of safety are subjective or qualitative Preferred techniques for quantifying safety remain debated Chosen safety indicators influence: Identification of hazardous sites Selection of countermeasures Reported impacts of treatments Studies have typically relied on actual crash frequency and severity Crash-based methods are retrospective and require long collection periods
5/32 Introduction Traditional Methods
6/32 Introduction Surrogate Safety Methods In response, surrogate safety measures have become popular Any non-crash measures that are physically and predictably related to crashes and can be used to estimate collision risk Conflict techniques rely on the observation of road user interactions Most commonly measured using time-to-collision (TTC) or postencroachment time (PET) Challenges and limitations of conflict techniques: If measured manually, cost and subjectivity If automated using video footage, the process is resource intensive and may be sensitive to potential speed bias in video data
7/32 Introduction Surrogate Safety Methods Other techniques rely on traffic parameters such as volume and speed Traffic parameters share a complex relationship with crash occurrence Maintaining an adequate network of permanent infrastructure is impractical Loop detectors have typically led to reliance on aggregate traffic data Non-intrusive sensors have assisted in overcoming these Evaluation of these sensors in reporting safety surrogates has been rare
8/32 Introduction Purpose and Objectives Purpose To evaluate how non-intrusive sensors (microwave radar, magnetometer, and video) report safety indicators based on traffic parameters Objectives Explore the use of sensors in collecting microscopic traffic data for computing surrogate safety measures Consider the usefulness of various traffic parameters as surrogate safety measures in urban environments Discuss potential implications of sensor technologies in safety research
9/32 Outline Introduction Literature Review Methodology Site Selection Technology Selection and Instrumentation Data Analysis Results Conflict Analysis Temporal Speed Variation Lateral Speed Variation Conclusions
10/32 Literature Review Theory Stable traffic conditions result in high levels of safety (Oh et al. 2001) Variation in traffic flow may be more important than average measures (Lee et al. 2002) Variation can be longitudinal, lateral, or temporal Evidence Researches have begun to understand the traffic flow/safety relationship the key elements of traffic flow affecting safety are not only mean volume and speed, but also variations in volume and speed (Golob et al. 2004) as individual vehicle speeds deviate [ ] from the average speed [ ] the probability of having a crash increases (Abdel-Aty and Pande 2005)
11/32 Literature Review Shortcomings Existing work has focused on loop detector data No attempt to compare surrogate measures reported by different technologies capable of providing vehicle-level data Previous studies have examined traffic flow on freeways Must consider other facility types in urban areas The use of aggregate data represents a potential for ecological fallacy arises whenever an observed statistical relationship between aggregated variables is falsely attributed to the units over which they were aggregated (Davis 2002)
12/32 Outline Introduction Literature Review Methodology Site Selection Technology Selection and Instrumentation Data Analysis Results Conflict Analysis Temporal Speed Variation Lateral Speed Variation Conclusions
13/32 Methodology Site Selection Site 1: University Avenue near Milton Street Local street in Montreal, Quebec One way, one traffic lane Posted speed limit of 30 km/h Instrumented for two hours Nearly 900 detection records
14/32 Methodology Site Selection Site 2: Taschereau Boulevard near Lapinière Boulevard Urban arterial in Brossard, Quebec Five lanes in two directions Posted speed limit of 50 km/h Instrumented for four hours Approximately 5000 detection records in the southbound direction
15/32 Methodology Technology Selection and Instrumentation Video data was captured using a commercially available camera Extracted using Traffic Intelligence, developed at Polytechnique Montreal Radar and camera were mounted to telescoping pole The video camera was mounted at 9 m, perpendicular to traffic The radar was mounted between 4 and 5 m, perpendicular to traffic Installed at Sites 1 and 2 The magnetometer was affixed directly to the pavement surface in the line-of-sight of the radar and camera Installed at Site 1 only
16/32 Methodology Technology Selection and Instrumentation
17/32 Methodology Conflict Analysis Rear-end TTC can be estimated using microscopic traffic data the time required for two vehicles to collide if they continue at their present speed and on the same path TTC 1,2 = V 1G 1,2 V 2 V 1 Vehicles experiencing a TTC under a certain threshold are said to engage in a more severe conflict (higher potential for collisions) In practice, the threshold has been set between 3 and 5 seconds Conflict analysis was conducted using data from Site 1 Speed and gap distributions were generated and compared
18/32 Methodology Temporal Speed Variation As variation in speed increases drivers have to adjust their speed more frequently and they are more likely to make misjudgement Temporal speed variation was evaluated using the Coefficient of Variation of Speed (CVS) N CVS i = 1 N n=1 (σs ) n,i s n,i CVS was calculated using data from Site 2 Manual ground truth sampling was provided from 8:30 AM to 9:00 AM
19/32 Methodology Lateral Speed Variation Lateral speed variation may indicate potential for sideswipe collisions Lateral speed variation was quantified using the average speed difference across adjacent lanes (ΔS) ΔS i = 1 N 1 N 1 n=1 s n,i s n,i+1 ΔS was calculated using data from Site 2 Speed profiles were generated for the corridor at each time period Manual ground truth sampling was provided from 8:30 AM to 9:00 AM
20/32 Outline Introduction Literature Review Methodology Site Selection Technology Selection and Instrumentation Data Analysis Results Conflict Analysis Temporal Speed Variation Lateral Speed Variation Conclusions
Cumulative Proportion Cumulative Proportion CARSP 2015 Collision Prediction and Prevention Approaches 21/32 Results Conflict Analysis Mean (Standard Deviation) Video Radar Magnetometer Gap (s) 6.6 (9.0) 8.6 (11.2) 8.5 (10.9) Speed (km/h) 37.1 (12.5) 29.4 (11.0) 32.6 (11.7) 1.0 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 1 3 5 7 9 11 13 15 17 19 Gap (seconds) 0.0 2 10 18 26 34 42 50 58 66 Speed (km/h)
Cumulative Frequency CARSP 2015 Collision Prediction and Prevention Approaches 22/32 Results Conflict Analysis 120 100 80 60 40 20 0 1 2 3 4 5 6 7 8 9 10 TTC (seconds)
23/32 Results Conflict Analysis Video measured smaller gap times Video counts more vehicles due to false detections and over-segmentation Gap times from radar and magnetometer were not statistically different Video measured significantly higher speeds Due to overestimation bias present in video-based sensors Speeds from radar and magnetometer were also significantly different The video reported more severe conflicts at every value of TTC The radar and magnetometer performed almost identically The radar and magnetometer often miss vehicle speed measurements, without which TTC computation is not possible
Coefficient of Variation of Speed Mean Speed (km/h) Standard Deviation of Speed (km/h) CARSP 2015 Collision Prediction and Prevention Approaches 24/32 Results Temporal Speed Variation 85 80 75 70 65 60 55 50 45 40 35 8:00 8:30 9:00 9:30 10:00 10:30 11:00 Time 40 35 30 25 20 15 10 5 0 8:00 8:30 9:00 9:30 10:00 10:30 11:00 Time 0.6 0.5 0.4 0.3 0.2 0.1 0.0 8:00 8:30 9:00 9:30 10:00 10:30 11:00 Time
25/32 Results Temporal Speed Variation Speed overestimation bias is apparent in the video data The relative pattern of speeds collected by each detector is similar CVS was consistently higher for the radar than for the video The radar reports a lower level of safety compared to the video The equation for CVS contains average speed in the denominator The overestimation bias in the video data decreases the reported risk Compared to the ground truth data: Both sensors showed good accuracy between 8:30 AM and 8:45 AM Both sensors consistently overestimated CVS from 8:45 AM to 9:00 AM
Speed Difference Across Lanes (km/h) Mean Speed (km/h) Mean Speed (km/h) CARSP 2015 Collision Prediction and Prevention Approaches 26/32 Results Lateral Speed Variation 80 80 75 75 70 70 65 65 60 60 55 55 50 50 45 45 40 40 35 35 30 1 2 3 4 5 30 1 2 3 4 5 Lane Lane 14 12 10 8 6 4 2 0 8:00 8:30 9:00 9:30 10:00 10:30 11:00 Time
27/32 Results Lateral Speed Variation As this measure uses only the difference in measured speed across lanes, the sensors perform similarly The overestimation bias of the video data is inconsequential There was no statistical difference in the mean value of ΔS reported by video and radar Compared to the ground truth data: The video performed more accurately from 8:30 AM to 8:45 AM The radar performed more accurately from 8:45 AM to 9:00 AM
28/32 Results TTC Video Radar Magnetometer # <3 seconds 30 10 9 # <5 Seconds 51 23 16 CVS % Highest 19% 81% - Highest Value 0.34 0.48 - ΔS % Highest 57% 43% - Highest Value 13.21 10.49 -
29/32 Outline Introduction Literature Review Methodology Site Selection Technology Selection and Instrumentation Data Analysis Results Conflict Analysis Temporal Speed Variation Lateral Speed Variation Conclusions
30/32 Conclusions For rear-end TTC, the video-based sensor found more severe conflicts Therefore, a greater estimated risk compared to the other sensors CVS indicated by the radar was consistently higher than for the video Video data should only be used if the overestimation bias can be removed Utilizing ΔS, the results from both detectors are similar as expected Failed to show a statistical difference in the results of both detectors Variability in speed data leads to uncertainty for all sensors Speed measurement is less precise compared to manual methods (residual error is larger for automated methods)
31/32 Conclusions Future Work Accurately quantifying safety requires selecting both an appropriate indicator and an appropriate device An ideal method should consider multiple indicators to improve redundancy The computed surrogate measures need additional consideration Work is needed to determine TTC thresholds for urban arterials Is TTC threshold enough? Consider the use of other data collection techniques, including GPS data from probe vehicles Correlation between indicators and actual risk must be determined
32/32 Thank you! Questions or comments? joshua.stipancic@mail.mcgill.ca