Crash-Based Safety Testing of Autonomous Vehicles: Insights From Generating Safety-Critical Scenarios Based on In-Depth Crash Data

Safety is a paramount concern in the development of Autonomous Vehicles (AVs) within the automotive industry. Ensuring AVs’ reliable performance requires the development of varied and realistic testing scenarios. Identifying unknown hazardous scenarios presents a notable challenge in the testing process. To address a wider spectrum of unknown hazardous scenarios, this study introduces a novel methodology for generating safety-critical scenarios for AV safety testing based on in-depth crash data. First, we used the China In-depth Mobility Safety Study-Traffic Accident (CIMSS-TA) database to decompose crash scenarios into static and dynamic variable tuples. Subsequently, the time series Generative Adversarial Networks (TimeGAN) was used to capture temporal features and to generate safety-critical scenarios using the original dataset. The synthesized scenarios were then assessed in terms of similarity, usability, and risk. To further validate the proposed method, we compared it with other generative models, highlighting the superior performance of TimeGAN. Finally, a well-known autonomous driving system, Baidu Apollo, was tested in original, synthesized, and naturalistic driving scenarios. Experimental results demonstrate that safety-critical scenarios generated from in-depth crash data preserve essential risk characteristics while increasing diversity. Baidu Apollo demonstrated superior safety performance in hazardous scenarios compared with human drivers, while naturalistic driving scenarios were less efficient in evaluating the performance of autonomous driving systems in safety-critical events. The synthesized scenarios provide a comprehensive dataset for testing AVs under high-risk conditions, overcoming the limitations of crash data and expanding the scope of safety-critical scenarios to improve AV technology evaluation and innovation.