Socially-Aware Graph Convolutional Network for Human Trajectory Prediction

Learning to understand human behaviors and predict their trajectories is a prerequisite for an automated car to navigate through the crowd safely and efficiently. This problem is particularly challenging as it requires the car to jointly reason about multiple pedestrians in a scenario where every one cooperates with each other to avoid collisions. To model the interactions among them, we propose a socially-aware graph convolutional network (SAGCN) which solves this problem in a graph learning framework. An attention graph is first built where each of its node carries the pedestrian temporal information and its edge represents the correspondence between pairwise pedestrians. To extract the temporal features, we implement temporal convolutional network (TCN) on each node. By utilization of relative motion between pairwise pedestrians, another TCN is employed to learn the correspondence of them which is formulated to an adjacency matrix of the attention graph. With the learned temporal features and adjacency matrix, a graph convolutional network (GCN) is exploited to aggregate the node information and jointly predict future trajectories of multiple pedestrians. Through experiments in several publicly available datasets, we demonstrate that our model effectively learns the comprehensive spatial-temporal representation and outperforms state-of-art methods in terms of prediction accuracy.