PFM-TMLC: Physics-Guided microscopic trajectory reconstruction from fixed sensors under emergencies

Urban traffic monitoring systems often suffer from trajectory loss due to occlusion, blind spots, and sensing failures, which become particularly critical under emergencies such as incidents, evacuations, or extreme weather. In these conditions, incomplete trajectories can degrade situational awareness, bias traffic state estimation, and delay operational responses, thereby undermining the resilience of smart city transport systems. This study proposes the Particle-Filtered Mamba − Trajectory Matching and Lane Change recognition (PFM-TMLC) model, a unified microscopic trajectory reconstruction framework that integrates three complementary modules. First, a Mamba-based longitudinal reconstruction model learns long-range temporal dependencies from fixed-sensor observations and surrounding-vehicle states, enabling accurate recovery of missing longitudinal motion. Second, a particle-filter refinement model enforces physics consistent evolution and scenario adaptive constraints to filter and correct data-driven predictions, improving robustness in different conditions while preventing physically implausible behaviors. Third, a multi-task learning component jointly performs trajectory matching and lane-change recognition, allowing disjoint trajectories to be reconciled at the trajectory level and lane transitions to be inferred during missing intervals. Experiments on the NGSIM I-80 dataset across four representative car-following scenarios show that the particle-filter refinement reduces longitudinal reconstruction errors substantially relative to the baseline, achieving up to about 47 % reduction in position and speed Root Mean Squared Error (RMSE), while eliminating collision cases observed in purely data-driven baselines. The multi-task module further delivers reliable lateral completion, with trajectory matching accuracy exceeding 82 % and lane-change direction accuracy exceeding 83 % for left and 72 % for right maneuvers. These results indicate that PFM-TMLC model can provide accurate and physically plausible trajectory completion to support robust traffic operations, especially when sensing quality deteriorates during emergencies.