Pedestrian trajectory prediction based on improved avoidance force algorithm

Abstract The complexity of interactions between pedestrians poses a challenge to pedestrian trajectory prediction, and existing trajectory prediction methods based on data-driven models lack interpretation for modeling interactions between pedestrians. To address this problem, an improved avoidance force algorithm is proposed to model the interaction of pedestrian forces explicitly. Multiple socially acceptable pedestrian trajectory information is generated by using the prior knowledge of observed trajectory and the avoidance force algorithm.The avoidance force trajectories are evaluated by an attention network to generate confidence scores; the avoidance force trajectories are selected based on the confidence scores; The final accurate trajectories are refined using Teacher-forcing. In comparison with recent approaches, our experimental results on the ETH and UCY datasets demonstrate a significant improvement in both Average Displacement Error (ADE) and Final Displacement Error (FDE) achieved by the proposed method.