Evaluation of Pedestrian Future Conflict at Non-Signalized Intersections Based on Probabilistic Trajectory Prediction

Recent advances in sensing and artificial intelligence technologies have laid the foundation for proactive pedestrian safety at non-signalized intersections. These technologies enable the prediction of future pedestrian-vehicle conflicts, allowing for timely intervention. Existing studies, however, overlook the uncertainty inherent in future prediction, which mainly stems from behavioral variability and observational inaccuracies. Probabilistic trajectory prediction can capture these uncertainties, yet it has rarely been applied to future conflict evaluation. To improve future conflict evaluation by accounting for uncertainty, this study introduces a novel method that integrates probabilistic trajectory prediction with a newly designed surrogate safety measure, Stochastic Predicted Time to Collision (SPTTC). SPTTC enables accurate conflict prediction, reliable scenario classification and evaluation of conflict severity progression, as demonstrated in quantitative results and case studies. Across different types of non-signalized intersections, SPTTC was separately integrated with four probabilistic trajectory prediction models, achieving F1 scores of 0.77–0.84 and consistently outperforming baseline methods. The proposed method also exhibits strong robustness under high-uncertainty conditions, such as unpredictable pedestrian behavior (e.g., children), low visibility (e.g., nighttime), and missing data. This proposed method can be integrated into proactive pedestrian protection systems, enhancing safety by providing timely alerts to pedestrians or activating vehicle brakes at critical moments.