A Defect Scenario Generalization Generation Method for Autonomous Driving System Testing

A defect scene is a collection of scenes with a high defect triggering rate of an autonomous driving system, which has the characteristic of being distributed in the logical scene parameter space between the collision hazard and the safety boundary. In order to solve the problem of low defect scene coverage and scene complexity in the test scenarios of autonomous driving systems, a generalized defect scene generation method based on the attention mechanism is proposed to achieve the automatic generation of high-complexity defect scenes. Firstly, the scene is abstracted into a logical scene model that changes dynamically over time by combining environmental information and vehicle dynamics model; then, the scene controller modeling problem is abstracted into a Markov decision process, and the SA-DDQN (Social Attention Double Deep Q-Network) algorithm is proposed based on the attention mechanism. Constructing the multi-attention mechanism into the strategy model can make the ego vehicle more effective in capturing the complex interactions between input variables and focusing on potentially dangerous vehicles that are close to or in conflict with the planned path. Thus, it solves the problems of low percentage of defective scenarios, low generation efficiency, low scenario complexity, etc., and accelerates the testing process. Finally, a typical highway lane-changing overtaking scenario is taken as an example for simulation verification. The results show that the defective scene model generated based on the proposed SA-DDQN algorithm performs well in terms of scene interaction, scene coverage and repeatable testing.