An Efficient High-Risk Lane-Changing Scenario Edge Cases Generation Method for Autonomous Vehicle Safety Testing

Safety-critical scenarios for autonomous vehicle testing have been rare in the real world. However, virtual simulation technology can provide numerous and diverse testing scenarios. Various effective methods can be employed to generate unknown safety scenarios and determine dangerous scenarios for tested autonomous vehicles. To meet time-varying safety-critical scenario construction requirements for virtual testing of autonomous vehicles, this study proposes a data-model-driven method for generating the edge cases of high-risk lane-changing scenarios. The trajectory time generative adversarial network named the Traj-TimeGAN is proposed to generate emergency lane-changing trajectories based on data from HighD dataset. In addition, a safety distance-based constraint model is designed to define the safety boundary and generate the initial state of a tested autonomous vehicle. Further, a generalization generation method is developed to generate many risk-critical scenario edge cases, and a lane-changing scenario dataset is constructed. The proposed method is verified by experiments, and the average root-mean-square error for 520,000 generated emergency trajectories is
$1.9 \times 10^{- 3}$
, indicating a high similarity between generated and real trajectories. Based on the results, in 99.93% of the 520,000 generated risk-critical lane-changing edge cases, the absolute value of time to collision (TTC) between the tested automated vehicle and the lane-changing background vehicle is less than 1 s. Finally, the results show that the proposed method can effectively generate high-risk lane-changing edge cases for autonomous vehicle testing.