CRF Based Pedestrian Trajectory Reconstruction With Spatio-Temporal Feature Embedding and Entropy Constraints

Pedestrian trajectory reconstruction is crucial for understanding urban activity patterns, particularly in crime prevention and investigation. Advances in re-identification technology enable the association of pedestrians across extensive video, making video trajectories a vital data source for trajectory mining. Our objective is to develop an effective method for reconstructing pedestrian paths from video trajectories. Existing methods face three main issues: 1) they fail to account for data scale inconsistencies between trajectories and road networks, limiting candidate path diversity; 2) they do not fully integrate distance, direction, and surveillance coverage, leading to inferred paths that do not accurately represent movement characteristics; 3) they do not effectively constrain uncertainty within trajectory data, causing it to propagate during reconstruction. To address these issues, we propose a method based on conditional random field with spatio-temporal feature embedding under entropy constraints (CRF-STEEC). This method standardizes both trajectory and road network data to a uniform scale using adaptive gridding and generates various likely paths. It then infers paths by evaluating their similarity to the target’s spatiotemporal movement characteristics, involving the extraction and embedding of spatiotemporal state features. Finally, to control uncertainty, it analyzes transition features based on trajectory entropy constraints to prevent the accumulation of uncertainty. Experimental results show that CRF-STEEC significantly improves the robustness and accuracy of trajectory reconstruction. Our source code is available at: https://github.com/lee1906068/CRF_STEEC.