From real-world trajectory-level pre-crash data to driving behavior modeling: An inverse reinforcement learning framework based on ghost-probe scenarios

The validity of autonomous vehicle testing depends on the construction of scenarios that exhibit human-like driving behaviors. Most existing studies rely on naturalistic driving data to model driving behaviors and replicate real-world traffic scenarios, while research based on crash data has largely remained at a qualitative level, focusing on causation analysis without modeling driving behaviors under crash scenarios. To bridge this gap, this study introduces the PRe-crash Interaction Scenario Modeling framework (PRISM), a complete framework that connects the collection of in-depth real-world pre-crash data with the modeling of pre-crash driving behaviors. In this work, PRISM is instantiated primarily using trajectory-level pre-crash interactions extracted from “ghost-probe” crash data. Within PRISM, inverse reinforcement learning (IRL) is employed to recover the reward functions underlying drivers’ decision-making prior to crashes, thereby uncovering how drivers react in the moments leading up to a crash and supporting the construction of realistic pre-crash scenarios. Finally, experimental comparisons between maximum entropy IRL (MEIRL) and maximum entropy deep IRL (MEDIRL) demonstrate that MEDIRL, by capturing nonlinear behavioral patterns, more faithfully reconstructs driving behaviors before crashes. The comparison of generated scenarios with real crash scenarios across multiple similarity metrics and surrogate safety metrics confirms the validity of PRISM. This validation highlights PRISM’s potential to advance the generation of human-like scenarios and strengthen the reliability of autonomous vehicle testing.