Multivehicle cooperative trajectory prediction at urban intersections based on dual-stream temporal-spatial attention fusion

Traffic status in urban intersections is highly dynamic and the interaction among multiple vehicles is complicated, which make precise trajectory prediction in urban intersections challenging. To tackle this problem, this paper introduces a new dual-stream temporal-spatial attention fusion network (DTSAF-Net). The model proposes a double-branch attention mechanism to capture the temporal dynamic features, and spatial interaction features of the vehicle, and achieves mutual transformation in the joint representation space. More precisely, the time series block leverages multi-layer stacked multihead Transformer encoders to model the long-range time-dependent information on the historical trajectory of a bicycle. The spatial branch builds an interaction graph that is dependent on a dynamic neighborhood of each vehicle and applies a graph attention mechanism to learn real-time interaction among various vehicle at the current frame. The two branches are fused by a trans-attentional fusion module with adaptively concurrent evolutionary and dynamic dependences. And a multi-scale trajectory refinement decoder is introduced for enhancing the prediction accuracy and time-domain consistency by optimizing the prediction trajectory for different time periods in the future from the top to bottom layer by layer. The devised DTSAF-Net obtains better performance than other the state-of-the-art approaches in both Mean Displacement Error (ADE) and Final Displacement Error (FDE), especially in high-density interaction scenes.