Predicted lane-based time to collision: A novel surrogate safety measure for pedestrian potential risk evaluation at non-signalized intersections

Pedestrian crossing safety at non-signalized intersections remains a critical challenge. Proactive pedestrian protection systems, such as pedestrian warning systems, have emerged as reliable alternatives to crash data-dependent analyses. These systems use real-time sensor data to evaluate potential risk and enable timely interventions. Prediction-based Surrogate Safety Measures (SSMs), which project current states to estimate future conflicts, are widely adopted for risk evaluation. However, due to the complexity of pedestrian behavior, existing prediction-based SSMs are limited in accurately estimating conflict locations and arrival times, making long-distance risk evaluation particularly challenging. To address these limitations, this study proposes Predicted Lane-based Time to Collision (PLTTC), a novel surrogate safety measure for evaluating pedestrian risks at non-signalized intersections. PLTTC extends Lane-based PET by shifting from precise conflict points to a broader area-based approach and removes rigid kinematic assumptions by employing deep learning for arrival time prediction. Experiments with real-world CCTV data show that: (1) PLTTC outperforms existing SSMs in risk evaluation, particularly in long-distance scenarios, achieving an AUC of 0.92 on the Precision-Recall curve; (2) PLTTC demonstrates superior performance for road users with high variability in velocity, such as kids; and (3) PLTTC shows generalizability through cross-site validation, maintaining consistent performance across different intersection. By addressing long-distance scenarios, capturing high-variability behaviors, and demonstrating generalizability, PLTTC can enhance pedestrian warning systems and improves safety at non-signalized intersections. • Predicted Lane-based Time to Collision (PLTTC) is proposed as a new Surrogate Safety Measure. • PLTTC broadens conflict zones and applies deep learning for arrival time prediction. • Comparative results show improved risk identification in both short- and long-distance scenarios. • It provides higher accuracy for children with complex and variable crossing behaviors. • Validation at two Korean sites, day and night, confirms its robustness and applicability.