Pedestrian trajectory prediction method based on social force – Dynamic risk field coupled graph attention network

Abstract

Accurate pedestrian trajectory prediction plays a critical role in enhancing traffic safety at unsignalized intersections and advancing the deployment of autonomous driving technologies. To address the limitation of existing models in fully capturing the complex pedestrian-vehicle interactions at such intersections, this paper proposes a pedestrian trajectory prediction method based on a dual-domain coupling graph attention network that integrates social force and dynamic risk field models. The method employs an improved social force model to characterize pedestrian-to-pedestrian interactions and a dynamic risk field model to describe pedestrian-vehicle interactions. These interaction representations are mapped to the edge weights of the graph attention network, enabling adaptive fusion of multi-modal interaction effects. Furthermore, residual connections and a dynamic gating mechanism are incorporated to enhance feature propagation and adaptively balance the contributions of pedestrian and vehicle features. Finally, a LSTM-based encoder-decoder framework is utilized to generate the predicted trajectories. Experimental results on the DUT (Dalian University of Technology Anti-UAV Dataset) and SDD (Stanford Drone Dataset) demonstrate that the proposed method significantly improves the accuracy and reliability of pedestrian trajectory prediction in complex pedestrian-vehicle interaction scenarios.