STUGCN:A Social Spatio-Temporal Unifying Graph Convolutional Network for Trajectory Prediction

Abstract

Trajectory prediction, also known as trajectory forecasting, of interacting agents in dynamic scenes is a critical problem for many applications, including robotic systems and autonomous driving. Because of the complex interactions between the pedestrian, the problem poses a significant challenge. To predict future pedestrian trajectories, we propose a Spatio-Temporal Unifying Graph Convolutional Network (STUGCN) based on a Spatio-Temporal Graph Convolutional Network architecture. At each time step, the Spatio-temporal interactions captured by the Cross-Spacetime Skip Connections. Finally, in the temporal dimension of the aggregated features, a Time-extrapolator Convolutional Neural Network (TXP-CNN) is used to predict the pedestrians' future trajectories. In comparison to state-of-the-art methods, our model outperforms them on two publicly available crowd datasets (ETH and UCY) and achieves state-of-the-art performance.