Trajectory Prediction Method with Map Nodes Search for Intelligent Vehicles

Accurately inferring the future motion of neighboring vehicles is an indispensable capability for safe driving of intelligent vehicles. High-definition (HD) maps containing scene constraint information can dramatically improve the performance of trajectory prediction methods. In this paper, a prediction method via map nodes search is proposed to precisely predict the future movement of vehicles. The travelable path search strategy is proposed to search for travelable map nodes in a map that incorporates traffic flow information after graph attention (GAT) aggregation, and the set of map nodes obtained from the search is employed to guide the generation of predicted trajectories. In addition, the designed goal node search mechanism finds the goal node based on map nodes and considers it as a reference for the end point of the vehicle's future movement. Experimental results on nuScenes dataset illustrate that our prediction method is superior to baseline methods. Moreover, the map nodes search strategy can greatly improve the trajectory prediction accuracy and ensure the rationality of the predicted trajectory.