High-Risk Trajectories Generation for Safety Testing of Autonomous Vehicles Based on In-Depth Crash Data

Scenario-based testing offers a robust method for evaluating the safety of Autonomous Vehicles (AVs). In this study, we introduce LSTM-CGAN to generate scenarios for assessing AV safety performance. First, we extract static, environmental, and dynamic elements from the China In-depth Mobility Safety Study-Traffic Accident (CIMSS-TA) dataset and reconstruct 630 real-world crashes to obtain pre-crash trajectories of the ego and objective vehicles. The proposed model then synthesizes samples by utilizing high-risk trajectories extracted from pre-crash trajectories, resulting in the generation of 6,300 trajectories, ten times the original dataset. The synthesized samples closely resemble real-world trajectories while enhancing diversity, with 61.9% of road segment and 55.5% of intersection trajectories scenes being classified as challenging. By integrating static and dynamic elements with the synthesized trajectories, critical scenarios for counterfactual simulations are generated. Finally, the Baidu Apollo, a well-known Autonomous Driving System (ADS), is tested in simulations using both generated and original scenarios. It is evidenced that the proposed method is effective in generating more safety-critical scenarios and that static and environmental factors have different impacts on AV performance. This study provides valuable insights into the unknown safety-critical scenario construction of AV safety testing, thereby facilitating the commercialization of AVs.