Rapid Generation of Hazardous Multi-Participants Scenarios for Testing Intelligent Driving System

In the process of autonomous development, potential defects should be discovered and solved to ensure the Safety of the Intended Functionality (SOTIF), and the scenario-based test is currently the mainstream approach of testing autonomous. However, most existing scenario generation methods either reduce the behavioral details of traffic participants in parameterization, or take expensive time consumption, especially when generating the scenarios involving multiple traffic participants in real-time tests. In this paper, we propose a generation framework involving a cost-based controller and a propagating knowledge library, which can generate hazardous multi-participants scenarios and form sectionalized complex behavior to reflect diverse failure forms. We carry out a real-time test with the system under test (SUT) in the loop to verify our proposed framework, and the results show that our framework can generate expected scenario cases for a black-box target within several minutes, and the generation results of failure forms are better-distributed compared to parameterizing methods.