Multi-scale trajectory reconstruction for freeway traffic via deep reinforcement learning under heterogeneous data

Abstract

Precise vehicle trajectory data is essential for traffic flow modeling, trajectory prediction, and energy consumption estimation. However, fixed detectors yield only sparse point-based observations, while mobile detectors such as probe vehicles (PVs) provide complete but low-frequency trajectories, making it difficult to directly capture full vehicle trajectories. To address this challenge, this study proposes a multi-scale trajectory reconstruction framework that focuses on lane-level spatiotemporal trajectories, leveraging macroscopic traffic states to guide the reconstruction of microscopic vehicle trajectories via deep reinforcement learning (DRL). First, an improved adaptive smoothing algorithm is developed to address data imbalance between fixed and mobile detectors, constructing a macroscopic velocity field that serves as both the decision environment and the reward reference for the DRL agent. Second, based on the two-dimensional intelligent driver model (2D-IDM) and its extended version, a set of bidirectional candidate trajectories incorporating driver stochasticity is generated by jointly considering the upstream and downstream PV behaviors, providing physically plausible microscopic priors. The DRL agent then learns an optimal trajectory fusion policy by minimizing the deviation between the fused velocity and the macroscopic field. The proposed framework is evaluated on NGSIM dataset under both free-flow and congested conditions. Experimental results show that the proposed method reduces speed errors by over 30.97 % and position errors by more than 20.12 % compared to baseline models, consistently achieving superior accuracy, stability, and generalization.