DSTAnet: A Trajectory Distribution-Aware Spatio-Temporal Attention Network for Vehicle Trajectory Prediction

Abstract

Trajectory prediction is a crucial task of autonomous driving and benefits vehicles travel safely in complex traffic environments. However, most existing trajectory prediction methods suffer from low accuracy issue due to the highly complex and dynamic interactions between vehicles, as well as the uncertainty of driver intentions in real-world scenarios. To address these challenges, we propose a trajectory Distribution-aware Spatio-Temporal Attention network (DSTAnet) for vehicle trajectory prediction. The proposed DSTAnet begins with a temporal feature extractor that combines Transformer Encoder Layers (TEL) and Gated Recurrent Units (GRU), extracting both long-term teporal dependencies and short-term variations in vehicle trajectories. Next, Graph Attention Networks (GAT) are introduced to extract spatial features in dynamically changing interactions, and focus on the most noteworthy vehicles, which are beneficial for modeling interactions between vehicles. Meanwhile, Mixed Density Network (MDN) is used to learn potential trajectory distributions from spatiotemporal features, which enriches the features related to driver intentions, providing a probabilistic understanding of possible future trajectories. By predicting multiple trajectories through random sampling, DSTAnet captures the inherent uncertainty of driver intentions. Finally, the INTERACTION dataset is adopted to estimate the model performance and results indicate that the proposed method outperforms others.