Traffic Flow Prediction with Vehicle Trajectories

This paper proposes a spatiotemporal deep learning framework, Trajectory-based Graph Neural Network (TrGNN), that mines the underlying causality of flows from historical vehicle trajectories and incorporates into road traffic prediction. The trajectory transition patterns are studied to explicitly model spatial demand via graph propagation along network; an attention mechanism is designed learn temporal dependencies based on neighborhood status; finally, fusion multi-step prediction integrated neural network design. proposed approach evaluated with real-world dataset. Experiment results show TrGNN achieves over 5% error reduction when compared state-of-the-art approaches across all metrics for normal traffic, up 14% atypical during peak hours or abnormal events. advantage transitions especially manifest itself in inferring high fluctuation as well non-recurrent flow patterns.