Spatial-Temporal Graph Neural Network For Interaction-Aware Vehicle Trajectory Prediction

In this paper, a Spatial Temporal Graph Neural Network (STGNN) model is developed, including a temporal block and Graph Neural Network (GNN) block, to solve the problem of vehicle trajectory prediction in unstructured scenes. Specifically, a temporal block combines a recurrent neural network and non-local operation to extract the features from past trajectories, and a GNN block models the subtle interactions between vehicles. The proposed model is evaluated on two datasets: Unstructured Scene Dataset and Argoverse Dataset. Experiment results show that the STGNN model achieves a better performance in the unstructured scenes and can be applied to common scenes where rules of the road dominate.