An Attention-Based Interaction-Aware Spatio-Temporal Graph Neural Network for Trajectory Prediction

Abstract

Pedestrian trajectory prediction in crowd scenes is very useful in many applications such as video surveillance, self-driving cars, and robotic systems; however, it remains a challenging task because of the complex interactions and uncertainties of crowd motions. In this paper, a novel trajectory prediction method called the Attention-based Interaction-aware Spatio-temporal Graph Neural Network (AST-GNN) is proposed. AST-GNN uses an Attention mechanism to capture the complex interactions among multiple pedestrians. The attention mechanism allows for a dynamic and adaptive summary of the interactions of the nearby pedestrians. When the attention matrix is obtained, it is formulated into a propagation matrix for graph neural networks. Finally, a Time-extrapolator Convolutional Neural Network (TXP-CNN) is used in the temporal dimension of the aggregated features to predict the future trajectories of the pedestrians. Experimental results on benchmark pedestrian datasets (ETH and UCY) reveal the competitive performances of AST-GNN in terms of both the final displace error (FDE) and average displacement error (ADE) as compared with state-of-the-art trajectory prediction methods.