Modeling crash avoidance behaviors in vehicle-pedestrian near-miss scenarios: Curvilinear time-to-collision and Mamba-driven deep reinforcement learning.
Journal:
Accident; analysis and prevention
PMID:
40043346
Abstract
Interactions between vehicle-pedestrian at intersections often lead to safety-critical situations. This study aims to model the crash avoidance behaviors of vehicles during interactions with pedestrians in near-miss scenarios, contributing to the development of collision avoidance systems and safety-aware traffic simulations. Unmanned aerial vehicles were leveraged to collect high-resolution trajectory data of vehicle-pedestrian at urban intersections. A new surrogate safety measure, curvilinear time-to-collision (CurvTTC), was employed to identify vehicle-pedestrian near-miss scenarios. CurvTTC takes into account the curved trajectories of road users instead of assuming straight-line future trajectories, making it particularly suitable for safety analysis at intersections, where turning vehicles usually follow curved paths. An effective algorithm considering predicted trajectories and collision types was designed to compute CurvTTC. When CurvTTC was applied to capture vehicle-pedestrian conflicts at intersections, it demonstrated superior performance in identifying risks more accurately compared to other surrogate safety measures, emphasizing the importance of considering the curved trajectories of road users. Further, a novel deep deterministic policy gradient based on the Mamba network (Mamba-DDPG) approach was used to model vehicles' crash avoidance behaviors during the vehicle-pedestrian conflicts captured. Results revealed that the Mamba-DDPG approach effectively learned the vehicle behaviors sequentially in both lateral and longitudinal dimensions during near-miss scenarios with pedestrians. The Mamba-DDPG approach achieved superior predictive accuracy by utilizing Mamba's dynamic data reweighting, which prioritizes critical states. This resulted in better performance compared to both the standard DDPG and the Transformer-enhanced DDPG (Transformer-DDPG) methods. The Mamba-DDPG approach was employed to reconstruct evasive trajectories of vehicles when approaching pedestrians and its effectiveness in capturing the underlying policy of crash avoidance behaviors was validated.