Using Behavioral Cloning for Rear-End Crash Scenario Generation

Abstract

Autonomous Driving (AD) technology has made significant breakthroughs in recent years. Virtual safety assessment is a primary method for evaluating the safety of the AD functions due to its efficiency, high experimental control, low risk, and costeffectiveness. However, the number of cases in existing crash datasets is insufficient to provide an adequate evaluation of critical scenarios. Therefore, it is necessary to generate synthetic crashes with high fidelity. Rear-end crashes, which has relatively low complexity (only involves the longitudinal movement of two vehicles) and high frequency (the most common crash type), is suitable as the first type for crash generation. Behavior Cloning (BC), which can reproduce vehicle behavior using a simple neural network structure, has the potential to generate synthetic crash scenarios with high fidelity. This thesis proposes a crash generation framework that combines a BC-based crash generation model with Hierarchical Clustering-based post-processing to generate crashes with the same distribution as the given crash dataset. In addition, as there is currently no established methodology for the assessment of generated cases, evaluation methods that combine both general statistical metrics and domain knowledge are introduced and implemented. In general, the pro posed framework is capable of generating crash kinematics with high fidelity, as the generated cases could pass both visual inspection and statistical evaluation. Future work should focus on enhancing the robustness of the generation framework. The generated synthetic crashes could be employed either for virtual safety assessments of AD functions or rear-end crash kinematic analysis for research purposes.