Autonomous Driving Testing Scenario Generation Based on In-Depth Vehicle-to-powered Two-Wheeler Crash Data in China.
Xinghua Wang,Yong Peng,Tuo Xu,Qian Xu,Xianhui Wu,Guoliang Xiang,Shengen Yi,Honggang Wang
DOI: https://doi.org/10.1016/j.aap.2022.106812
IF: 5.9
2022-01-01
Accident Analysis & Prevention
Abstract:A reliable critical-scenario-based safety assessment of autonomous vehicles in China requires a thorough understanding of complex crash scenarios in Chinese background traffic. Based on actual crashes between a vehicle and a powered two-wheeler (PTW) in China, this study generated the autonomous driving testing scenarios from functional, logical and concrete levels. First, 239 video-recorded crash cases were selected from the China In-depth mobility Safety Study - Traffic Accident (CIMSS-TA) database. Using the k-medoids clustering method, six functional scenarios were generalized according to seven crash characteristics (time of day, road type, road surface, obstruction, motion of vehicle, motion of PTW, relative moving direction and position of PTW with respect to vehicle), which contained two straight road scenarios, two T-junction scenarios and two intersection scenarios. Then, using a trajectory analysis program written by Python, the dangerous time instant of each crash was extracted based on the relative trajectory. According to five dynamic parameters of dangerous time instant, namely vehicle velocity (Vehicle_V), PTW X'-coordinate velocity (PTW_VX'), PTW Y'-coordinate velocity (PTW_VY'), PTW X'-coordinate relative position (PTW_LocX') and PTW Y'-coordinate relative position (PTW_LocY'), a crash trigger scheme was built to remain a case challenging when the involved vehicle is replaced by an autonomous vehicle with completely different maneuvers. Using the kernel density estimation (KDE), the logical scenarios were evolved by calculating the distribution of these dynamic parameters in each cluster. The results showed that there were differences in the distribution of dynamic parameters between six functional scenarios. For instance, the Vehicle_V in the scenario where a vehicle turning right impacts with a right/right rear PTW traveling straight ahead was higher than that in the scenario where a vehicle changing to the left lane impacts with a left/left rear PTW traveling straight ahead, with ranges of (10 km/h, 30 km/h) and (5 km/h, 15 km/h), respectively. Finally, considering the correlation of dynamic parameters, a virtual crash generation approach based on the independent component analysis (ICA) representing the original crashes with independent parameters was proposed to obtain sufficient concrete testing scenarios. The results showed that the statistical characteristics of virtual crashes were consistent with those of original crashes. Therefore, the virtual crash generation approach was effective. And a concrete crossing testing scenario with the crash trigger conditions of Vehicle_V = 26.272 km/h, PTW_VX' = 15.567 km/h, PTW_VY' = -1.670 km/h, PTW_LocX' = -27.265 m and PTW_LocY' = 52. 149 m was especially demonstrated. This study provides a theoretical basis for generating autonomous driving testing scenarios and data support for establishing relevant testing schemes tailored to the traffic environment in China.