Previous Abstract Return to Session D4 Next Abstract

Session D4: Ground Vehicle Navigation

Data-Driven Collision Risk Modeling for Connected and Automated Vehicles
J. Michael Wooten, Lakshay Narula, Matthew J. Murrian, Todd E. Humphreys, The University of Texas at Austin
Location: Spyglass

This paper presents a data-driven collision hazard model for quantification of Bayesian risk of collision: by fusing multiple connected vehicles to increase the number of aspects of opportunity on collision-able objects, account for the errors associated with vehicle-to-vehicle sensor sharing, and model detection probabilities of sensor modalities for specific object types. Through globally referenced shared sensing, a particle filter is used to estimate the probability of vehicle, pedestrian, and cyclist traffic in dynamic and high clutter environments.
The ubiquity of interactive info-tainment systems and impending deployment of wide scale automated vehicles increases the frequency of vehicle-pedestrian collisions due to inattention, missed detection, and/or nonmodeled hazards.
Non-occupant fatalities have continued to increase year over year and reached a record high of 7079 in 2016 [1]. Modern advanced driver-assistance systems (ADAS) utilize multiple sensing modalities to detect pedestrian and automotive traffic; however, dynamic traffic environments and cluttered urban areas lead to perception gaps with a high likelihood of missed detection of collision-able objects. Similarly, sensors of a specific type will fail in different manners.
Visual light cameras for instance experience a line of sight horizon reduction due to poor lighting at night and/or reduction of visible range in inclement weather. Automotive radar is more robust to low-light and weather related effects, but the chance of detection for pedestrians and cyclists is reduced overall. Also, the cost and complexity of newer solid state LIDAR systems limits the ease of wide scale deployment.
Even in the limit of overlapping sensor modalities and failure modes, there is a non-zero risk associated with blind spots. Most dynamic environments will produce certain areas of sensor shadowing where a particular sensor cannot perceive (e.g. a large bus would occlude line of sight perception of visible light cameras). Prior work has shown that probabilistic models of collision risk can increase the safety of vehicular transportation [2]; however, the efforts do not leverage multiple view points from connected vehicles, only explore LIDAR and stereo vision based perception, notably omitting radar sensing, without accounting for adverse environmental conditions, and models motion trajectory for detected objects only. Similarly, trajectory modeling and
hazard map construction can provide feedback for ADAS decision making engines,
but prior work has limited effectiveness for non-detected objects and has not explored real time map update and generation from real world vehicle data [3].
Enhanced sensor sharing through high data rate low latency links can reduce sensor blind spots, and support real-time local area map generation.
Collision risk modeling can be effectively posed as a particle filtering problem. The mutation step of the particle filter allows for class specific diffusion properties through specific motion models for pedestrian, cyclist and vehicle traffic. The selection step of filter accounts for the complex detection probability associated with sensing modality, particle type, vehicle dynamics, and environmental conditions. The resulting particle distributions are then distilled through Bayesian risk calculation to a hazard or risk map. Iterations of mutation and selection provide a real-time measure of collision risk.
The framework of Bayesian risk can be used to empirically show benefits and reduced risk of collision due to uncertainty for connected vehicles, and by extension, areas of high danger can be identified. These risk maps can inform roadway construction and/or infrastructure based mitigation for non-reducible sensor blindspots. Through multiple vehicle fusion, the marginal gain of additional sensors and aspects would aide in selection, design, and interface requirements of shared sensors in advanced connected and automated vehicles. On a micro scale, the risk map informs the automated vehicles decision engine such that lower risk and modified trajectories can be planned in complex and changing environments.

[1] National Highway Traffic Safety Administration, Department of Transportation, "2016 Fatal Motor Vehicle Crashes: Overview" (DOT HS 812 456) https://crashstats.nhtsa.dot.gov/Api/Public/Publication/812456 October 2017
[2] C. Laugier et al., "Probabilistic Analysis of Dynamic Scenes and Collision Risks Assessment to Improve Driving Safety," in IEEE Intelligent Transportation Systems Magazine, vol. 3, no. 4, pp. 4-19, winter 2011.
[3] A. Møgelmose, M. M. Trivedi and T. B. Moeslund, "Trajectory analysis and prediction for improved pedestrian safety: Integrated framework and evaluations," 2015 IEEE Intelligent Vehicles Symposium (IV), Seoul, 2015, pp. 330-335.



Previous Abstract Return to Session D4 Next Abstract