Answers are grounded in this paper and cite their sources.
Camera-based personal protective equipment (PPE) monitoring is commonly scored as object detection, whereas safety decisions concern individual workers. We present an auditable evidence-to-decision protocol that retains candidate boxes; assigns each PPE item to at most one detected worker; quantifies competing ownership; and distinguishes between automatic decisions, review requests, and missed people. Two independent annotators and a third adjudicator supplied worker-state and ownership references for a prespecified 200-image source test sample, a separate 50-image overlap stress sample, and 133 person-containing external test images. Checkpoints from three prespecified seeds for the You Only Look Once (YOLO) models YOLO11n and YOLO26n and their relation settings were selected on sequence-disjoint inner validation and evaluated on a frozen test. Mean source-primary end-to-end compliant F1 scores (each the harmonic mean of precision and recall) were 0.9367 for YOLO11n and 0.9520 for YOLO26n, with corresponding worker-detection coverage of 0.936 and 0.945. Geometric ownership agreed with 463/466 primary source-box labels, including 119/122 in multi-worker images. At a fixed 20% detected-worker review quota, a frozen correctness ranker reduced the primary automatic error in all six model–seed runs; adding ownership margins gave no stable further gain. Frozen source models detected fewer than one fifth of external workers on average; external-only fine-tuning recovered target worker coverage but sharply reduced source coverage. These results provide an independently checked and reproducible worker-level audit protocol and identify where detection, ownership, and review require site-specific validation.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: