A Generation Method of Unknown Unsafe Scenarios for Autonomous Vehicles

How to narrow the region of unknown unsafe scenarios is the key to improving the performance of Safety of Intended Functionality for Autonomous Vehicles. However, the research of the generation and evaluation method for unknown unsafe scenarios is still insufficient. In this paper, a novel unknown unsafe scenario generation method and an evaluation method for the generation results are proposed. Firstly, we build a generation system consisting of an evolving system module with spatiotemporally continuous scenarios and a scenario processing module to identify and classify valuable fragments to construct a classified scenario library. Secondly, the generation results are evaluated from 5 dimensions including Validity, Complexity, Efficiency, Diversity and Discrimination. According to the experiment results, with appropriate Background Vehicle (BV) settings, the proposed method can generate explainable, complex, diverse, discriminative unknown unsafe scenarios efficiently. And the evolving system has better performance with DRL-based BVs compared to MOBIL and Nilsson BVs in all evaluation dimensions. Finally, the limitations of MOBIL and Nilsson BVs in the generation of unknown unsafe scenarios are analyzed.