Lane-Aware Dynamic Spatio-Temporal Transformer for Unimodal Trajectory Prediction

Accurately predicting the future motion of traffic participants is essential for autonomous driving in complex urban traffic scenes. Due to the frequent interactions and positioning uncertainties in the perception system, it is important to enable efficient spatial and temporal modeling of diverse entities and provide reliable distributions for future trajectories. In this work, we propose a novel Lane-aware Dynamic Spatio-temporal Transformer, called DSTformer, for effective and accurate vehicle trajectory prediction. The system aims to improve the robustness of autonomous vehicles by incorporating lane-aware scene constraints and utilizing both dynamic activities and scene context information. Specifically, we incorporate two masking methods to mitigate the impact of deviations and uncertainties from temporal and geographic spatial propagation views. Vehicles' trajectories that violate the expected lane-following rules are penalized via the Transformer-based modeling. Semantic neighborhood information is also employed via a graph masking method to capture the long-range spatial dependencies in traffic data. Additionally, the centerline waypoints based map structure is used to capture the vehicle-lane relations and enable multimodal trajectory prediction. Experimental results on the real-world dataset demonstrate that the proposed framework is resilient to imperfect data and achieves competitive performance with fewer parameters compared to current state-of-the-art prediction methods.