Spatial-Temporal Attention Networks for Vehicle Trajectory Prediction

Abstract

Predicting the future trajectory of vehicles is essential to the safety of autonomous driving. However, due to the uncertainty of the future behavior of vehicles and the complexity of interactions between vehicles, reasonable and accurate trajectory prediction is still one of the huge challenges faced by autonomous driving. In this paper, we present a Spatial-Temporal Attention Networks (STAN) for the prediction of the future trajectory of vehicles. STAN uses Transformer network to extract the historical trajectory features of vehicles, uses Graph Attention Network (GAT) to extract the spatial interactions features between vehicles, and captures the temporal correlations of interactions through Transformer. The experimental results on the NGSIM US-101 dataset show that our model has achieved competitive results compared with some existing works.