EMSIN: Enhanced Multistream Interaction Network for Vehicle Trajectory Prediction

Predicting the future trajectories of dynamic traffic actors is the Gordian knot for autonomous vehicles to achieve collision-free driving. Most existing works suffer from a gap in characterizing the evolving interactions of scenario components and ensuring the physical feasibility of predictions, particularly in highly heterogeneous scenarios. Therefore, we propose an enhanced multistream interaction network (EMSIN), which is devoted to providing accurate trajectory predictions. The EMSIN highlights several threads of high-level time-varying interactions, including agent-traffic semantic, self-trend, and agent-agent dependencies. A novelly designed trend-aware mechanism is developed to capture the self-trend interactions from different representation subspaces sufficiently. To model the spatial information of traffic agents and extract their evolutions, we present a dynamic adaptive graph convolutional network that extends previously predefined graph paradigms. The adaptive and dynamic graphs in the EMSIN are created using learnable node embeddings, allowing the model to discern interaction strengths without additional attention modules. Finally, all high-level feature spaces elaborating multistream interactions are fused to generate possible agent actions with corresponding confidence values. Comprehensive experiments conducted on both L5kit and nuScenes datasets demonstrate that the EMSIN surpasses its counterparts, boasting smaller prediction errors and faster inference times. This study also introduces a fuzzy-based metric to probe the physical feasibility of predicted trajectories, providing valuable insights into appraising the performance of various prediction models from the perspective of fuzziness.