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

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