Abstract
Automated personal protective equipment (PPE) monitoring can support construction-site safety, but practical deployment requires reproducible performance, reliable confidence estimates, and robustness across varying field acquisition conditions. This study evaluates a modular two-stage framework comprising YOLO26s person detection and MobileNetV3-based multi-label classification of six PPE attributes: helmets, high-visibility vests, gloves, safety boots, goggles, and face masks. The dataset contained 1807 construction-site images and 7580 person crops, partitioned by video recording to prevent temporal and location-based leakage. The same MobileNetV3 architecture was trained in three independent runs using different random seeds. Attribute-specific thresholds were selected on the validation set, and probability calibration was assessed using temperature scaling and Expected Calibration Error. On the test set, Average Precision was highest for vests (0.995) and boots (0.979), followed by gloves (0.908), helmets (0.892), goggles (0.715), and masks (0.695). Calibration improved substantially for boots and vests, slightly for goggles, negligibly for helmets, and unfavourably for gloves and masks. Distance effects were strongly attribute-dependent, with goggles degrading markedly at 30 m, while worker-density effects were non-monotonic. The findings demonstrate the importance of reproducibility, calibration, and operational robustness when evaluating construction-site PPE classifiers.