MEVAR: Mobility-Enhanced Vehicle Trajectory Reconstruction with Camera Sensing Network

Vehicle trajectory reconstruction is crucial for understanding mobility patterns and supporting downstream applications. Existing approaches for camera-sensing data mainly cluster images by visual similarity, with limited spatio-temporal constraints, making them highly sensitive to image quality. We propose MEVAR, a mobility-enhanced framework that integrates an instance-level visual module with a road-embedded spatio-temporal point process to model the joint probability of trajectory points. This relaxes reliance on appearance features and enforces trajectory-level consistency. Furthermore, we introduce a self-supervised iterative optimization scheme that alternately refines the mobility model and clustering results without labeled data. Experiments on two large-scale highway datasets demonstrate that MEVAR achieves superior accuracy and generalization compared with strong baselines.