CSG: Critical Scenario Generation from Real Traffic Accidents
Zhang Xinxin, Li Fei, Xiangbin Wu
Abstract
Autonomous driving (AD) is getting closer to our life, but the severe traffic accidents of autonomous vehicle (AV) happened in the past several years warn us that the safety of AVs is still a big challenge for the AD industry. Before volume production, the automotive industry and regulators must ensure the AV can deal with dangerous scenarios. Although road test is the most common method to test the performance and safety of an AV, it has some manifest disadvantages, e.g., highly risky and unrepeatable, low efficiency and lack of useful critical scenarios. Critical-scenario-based simulation can effectively address these problems and become an important complement to road test. In this paper, we present a novel approach to extract critical scenarios from real traffic accident videos and re-generate them in a simulator. We also introduce our integrated toolkit for scenario extraction and scenario test. With the toolkit, we can build a critical scenario library quickly and use it as a benchmark for AV safety assessment, among other purposes. On top of this, we further introduce our safety assessment criteria and scoring method.