Joint Multi-Agent Trajectory Prediction for Cooperative Vehicle Infrastructure System

In the context of Cooperative Vehicle Infrastructure System(CVIS) communication, predicting multi-agent trajectories from a global perspective is crucial. Accurate prediction of behaviors and trajectories can provide decision-making guid-ance for Connected and Automated Vehicles(CAVs). This paper proposes a scene-consistent multi-agent trajectory prediction model. In this work, the scene is firstly divided into several weakly correlated subgraphs. Within each sub graph, A revised LaneGCN is employed to encode agents and HD-Map, and the joint prediction task is decomposed into a series of marginal and conditional predictions by proposed directed interaction graph predictor, increasing the interpretability of future predictions. Experiments conducted on the INTERACTION and Soochow roadside dataset validate the effectiveness of the model.