Message-passing framework for multi-camera worker tracking in construction

Current computer vision–based tracking systems face challenges in reliably associating identities of construction workers due to similar attire, frequent occlusions, and complex multi-view movements, leading to fragmented trajectories and ID switches. This work proposes a multi-camera tracking framework that detects workers in individual camera views and integrates observations across cameras using re-identification and message-passing. A region-based re-identification model enhances feature extraction for occluded workers and those wearing similar gear, producing more discriminative representations. Data association leverages message-passing approach to combine localization, visual features, and motion cues for robust clustering and trajectory generation. Experiments achieve IDF1 scores of 68.30 (controlled) and 85.10 (outdoor) with MOTA scores of 79.7 and 79.2, respectively. Results on the CAMPUS benchmark demonstrate strong generalization and competitive performance, meeting operational requirements for field deployment. The approach highlights potential for broader multi-camera tracking in occluded industrial environments, supporting applications in safety and productivity monitoring.