An Interpretable Collision Risk Prediction Model for Rear-End Near-Crash Scenarios Using CatBoost and SHAP

Predicting the collision risk is crucial for active traffic safety, as accurate collision risk prediction can help to take appropriate evasive behavior to effectively prevent rear-end collisions. Most existing studies have rarely analyzed the comprehensive impact of driving behavior and the driving environment on collision risk under different types of evasive behavior. To address this issue, this study extracts evasive events from vehicle trajectory data on highways and combines surrogate measures to quantify collision risk, aiming to establish a collision risk prediction model using CatBoost (Categorical Boosting) and SHAP (SHapley Additive exPlanation) methods. The results show that the proposed prediction model performed well, with accuracies of 90.36% and 91.77%, respectively. Due to the advantages of SHAP, the results of feature relative importance and specific impact analysis indicate that the influencing factors for collision risk vary among different types of evasive behavior, and the impact of road environment features cannot be ignored. The proposed model has the potential for implementation in ADASs to enhance their active safety capabilities.