PSA-GRU: Modeling Person-Social Twin-Attention Based on GRU for Pedestrian Trajectory Prediction

Abstract

This paper addresses the path prediction problem of multiple pedestrians interacting in dynamic scene, which is a great challenge because of the complexity and subjectivity of pedestrian movements. To forecast the future trajectory of a pedestrian, it is necessary to consider pedestrian subjective intention and social interaction information. Previous methods ignore that important position nodes of historical trajectories can reflect the subjective intention of pedestrians in complex trajectories. In this work, we present Person-Social Twin-Attention network based on Gated Recurrent Unit (PSA-GRU), which can fully utilize important position nodes of personal historical trajectory and social interaction information between pedestrians. In our approach, the person-attention encoding module extracts the most salient parts of personal history trajectory and helps model learn where and how pedestrians will go. Meanwhile, the social-attention pool module contributes to keep distance among different pedestrians and simulates the dynamic interaction of all pedestrians in real scenes. PSA-GRU also takes advantage of Gated Recurrent Unit (GRU) to enable the model to improve the computational efficiency. Experimental results demonstrate that the prediction accuracy and efficiency of our proposed model are both greatly improved on UCY and ETH datasets.