Research on vehicle accident hazard scenario derivation based on improved AST

Traffic accident scenarios serve as one of the critical sources for autonomous driving simulation testing. However, scenarios directly generated from traffic accident data for testing autonomous driving safety suffer from insufficient hazard. This paper proposes a scenario derivation method that integrates a scene tree model constructed based on accident data with an improved adaptive stress test. By establishing the accident scene tree model, the search output yielded six categories of vehicle-to-vehicle conflict samples, eight categories of vehicle-pedestrian conflict samples, six categories of vehicle-non-motorized two/three-wheeler conflict samples, and six categories of vehicle-motorized two/three-wheeler conflict samples. Finally, the collision scenarios were derived using an adaptive stress testing algorithm, and the generated scenarios were evaluated in terms of rationality and hazard. The results show that the generation rates of vehicle-to-vehicle collision scenarios, vehicle-pedestrian collision scenarios, and vehicle-two-wheeler scenarios are 11.97%, 12.28%, and 13.38%, respectively. The method proposed in this paper enhances the hazard level of generated scenarios, which exceeds that of real collision scenarios. The research findings can provide references for constructing and deriving hazardous scenarios in current autonomous driving.