Driving Collision Risk Modeling with Trajectory Prediction of Obstacle Vehicles

In order to address critical challenges in intelligent driving systems, including insufficient dynamic interaction modeling, limited accuracy in multimodal trajectory prediction, and over-reliance on single physical metrics for collision risk quantification. A proactive collision risk assessment framework is proposed by integrating probabilistic quantification with multimodal trajectory prediction. For trajectory prediction, a hierarchical graph attention network is developed to capture dynamic environmental features through adaptive fusion of high-definition maps, lane geometries, and vehicle motion history. A sliding window-optimized decoder is introduced within the conventional two-stage prediction architecture to refine trajectory outputs. For risk assessment, a probabilistic collision quantification method is designed to calculate collision likelihood between ego and surrounding vehicles based on predicted trajectories. Results on the Argoverse dataset demonstrate state-of-the-art performance with minimum final displacement error (=0.785), average displacement error (=1.157), and miss rate (=0.126), achieving 1% and 15.1% error reduction in endpoint prediction compared to HiVT and LaneGCN respectively. simulation of urban mobility, SUMO simulations reveal 5% deviation between predicted and actual risks, with risk fluctuation amplitude reduced by 33.3% and 18.75% against time to collision (TTC) and dynamic safety index (DSI) methods. The proposed model shows enhanced stability in continuous driving scenarios (risk fluctuation=0.3) and demonstrates superior accuracy in forecasting potential collision risks through systematic integration of trajectory prediction and probabilistic analysis. These findings validate the framework's effectiveness in proactive safety warning for intelligent vehicles.