DSTNet: Dynamic Trajectory Prediction for Autonomous Vehicles via Spatio-Temporal Attention

Predicting the future trajectories of target agents is critical for the safe operation of autonomous vehicles in complex environments. Accurate trajectory prediction requires the extraction of both temporal and spatial information. To address the limitations of existing static attention methods, which are constrained by fixed historical ranges and spatial radii under limited computational resources, a novel trajectory prediction approach using dynamic spatio-temporal attention is proposed. DSTNet incorporates a multi-scale spatial sparse attention module and a multi-scale historical causal attention module, both aimed at dynamically encoding the spatio-temporal dependencies of the data. These modules are unified with multi-modal interactive attention in a three-attention spatio-temporal module. Additionally, DSTNet includes an anchor-based trajectory refinement module, which adaptively optimizes the coarse-grained trajectories generated by the backbone network. Compared with state-of-the-art methods, DSTNet demonstrates outstanding performance on the Argoverse 1 and Argoverse 2 motion prediction benchmarks, delivering accurate and stable future trajectories. Notably, the model has been successfully deployed on an autonomous vehicle.