Dynamic Trajectory Planning for Urban Traffic Using Deep Learning-Based Traffic Prediction

Abstract

Traditional trajectory planning systems in navigation applications generate routes based on the current traffic conditions at the time of departure. However, urban traffic dynamics are highly time varying, and routes that are initially optimal may become congested during the trip, leading to increased travel times and inefficient routing decisions. This paper proposes a dynamic trajectory planning framework that integrates short-term traffic prediction into the routing process using deep learning-based time series models. Recurrent neural networks and long short-term memory networks are employed to forecast future traffic densities, which are then embedded into a rolling-horizon dynamic search to continuously update edge costs and re-optimize the route in real time. Simulation results on a temporal graph-based urban road network demonstrate that the proposed approach significantly improves routing performance compared to static planning. In particular, the prediction-enhanced dynamic routing strategy achieves an average reduction in travel time of approximately 12–18% and improves traffic density prediction accuracy by up to 25% in terms of mean-squared error compared to baseline static and non-predictive methods. These results confirm that explicitly coupling traffic prediction with online trajectory replanning enables more adaptive and efficient navigation under time-varying traffic conditions.