Map Search-Based Vehicle Trajectory Prediction Conditions for Lane Lines With Heterogeneous Interaction in Complex Urban Traffic

The embedding of high-level traffic semantics has elevated the precision of vehicle trajectory prediction tasks to a new level. However, owing to the absence of feature-level integration, the information from high-definition maps is underutilized. To this end, a map search-based vehicle trajectory prediction method conditioned on lane segments is proposed in this article. The map is discretized into a graph, where nodes represent lane centerline segments. On this basis, the agent-to-agent, agent-to-map, and map-to-map modules are designed to depict heterogeneous interaction patterns involving vehicles and pedestrians. In addition, a goal node querying mechanism is introduced, which integrates vehicle motion, interaction, and traffic flow states and serves as prior information for trajectory prediction. Finally, a feasible path selection strategy is proposed, generating traffic rule-related prediction trajectories point by point, fully utilizing map information. The experimental results on the nuScenes dataset indicate that the proposed method achieves state-of-the-art prediction accuracy compared with advanced methods.