Vehicle Trajectory Prediction Method Through Integrating Dynamic Risk Perception

Existing vehicle trajectory prediction methods excel in accuracy for conventional scenarios but often neglect risk control in high-risk situations, risking system failures during critical interactions. To address this gap, this paper proposes a novel framework that integrates dynamic risk perception into the prediction process. Our framework introduces a dynamic risk quantification mechanism that integrates temporal urgency, measured by Post-Encroachment Time (PET), and spatial threat, depicted by the Driver Risk Field (DRF), into a comprehensive risk index. This index is integrated into an LSTM-based encoder-decoder model via a risk-integrated attention mechanism, guiding the model to focus on potentially hazardous moments. To overcome the scarcity of relevant data, we constructed a large-scale, high-risk trajectory dataset using Monte Carlo simulation. Experiments on this dataset demonstrate the superiority of our approach. Compared to strong baselines, our method not only maintains competitive geometric accuracy but also significantly improves key safety metrics. It reduces risk MSE by 12.4%, collision velocity MSE by 20.3%, and, most critically, lowers the collision miss rate by 53.3%. These results validate that our risk-aware framework produces substantially safer and more reliable trajectory predictions, effectively improving the safety assurance of autonomous vehicles in challenging scenarios.