AST-GNN: An attention-based spatio-temporal graph neural network for Interaction-aware pedestrian trajectory prediction

Predicting pedestrian trajectories in the future is a basic research topic in many real applications, such as video surveillance, self-driving cars, and robotic systems. There are two major challenges in this task, the complex interaction modeling among pedestrians and the unique motion pattern extraction for each pedestrian. Regarding the two challenges, an attention-based interaction-aware spatio-temporal graph neural network is proposed for predicting pedestrian trajectories. There are two components in the proposed method: spatial graph neural network for interaction modeling, and temporal graph neural network for motion feature extraction. Spatial graph neural network uses an attention mechanism to capture the spatial interactions among all the pedestrians at each time step. Meanwhile, temporal graph neural network uses an attention mechanism to capture the temporal motion pattern of each pedestrian. Finally, a time-extrapolator convolutional neural network is used in the temporal dimension of the aggregated graph features to predict the future trajectories. Experimental results on two benchmark pedestrian trajectory prediction datasets demonstrate the competitive performances of the proposed method in terms of both the final displace error and the average displacement error metrics as compared with state-of-the-art trajectory prediction methods.