Trajectory prediction for autonomous driving based on multiscale spatial-temporal graph

Predicting the trajectories of surrounding heterogeneous traffic agents is critical for the decision making of an autonomous vehicle. Recently, many existing prediction methods have focused on capturing interactions between agents to improve prediction accuracy. However, few methods pay attention to the temporal dependencies of interactions that there are different behavioural interactions at different time scales. In this work, the authors propose a novel framework for trajectory prediction by stacking spatial-temporal layers at multiple time scales. Firstly, the authors design three kinds of adjacency matrices to capture more genuine spatial dependencies rather than a fixed adjacency matrix. Then, a novel dilated temporal convolution is developed to handle temporal dependencies. Benefiting from the dilated temporal convolution, the authors' graph convolution is able to aggregate information from neighbours at different time scales by stacking spatial-temporal layers. Finally, a long short-term memory networks (LSTM)-based trajectory generation module is used to receive the features extracted by the spatial-temporal graph and generate the future trajectories for all observed traffic agents simultaneously. The authors evaluate the proposed model on the publicly available next generation simulation dataset (NGSIM), the highway drone dataset (highD), and ApolloScape datasets. The results demonstrate that the authors' approach achieves state-of-the-art performance. Furthermore, the proposed method ranked #1 on the leaderboard of the ApolloScape trajectory competition in March 2021.