A crash-injury-aware driving safety field for real-time risk assessment and its application in autonomous vehicle motion planning

Abstract

Driving risk comprises not only the likelihood of a collision but also the severity of its potential consequences. In this paper, we introduce a crash-injury-aware driving safety field (CIA-DSF) for real-time risk assessment, which explicitly incorporates key features that influence crash injury severity into the potential field formulation. First, an interpretable machine learning framework combining eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) is developed to identify the key features and qualitatively analyze their complex interactions in determining crash injury severity. Then, these extracted features are integrated into the driving safety field, accompanied by the introduction of impact area tuning factor to capture how vehicle geometry influences the spatial distribution of injury severity. Next, the CIA-DSF is utilized within a model predictive control (MPC)-based motion planner, enabling autonomous vehicles to proactively anticipate and effectively mitigate driving risks. Finally, three representative case studies are conducted to validate the effectiveness and superiority of the proposed risk assessment framework and the corresponding motion planner. Simulation results demonstrate that both the shape and intensity of the potential field are continuously updated in respond to collision probability and anticipated injury severity, and the motion planner consistently guides autonomous vehicles away from scenarios that could result in severe injuries if a collision were to occur.