A person-centric framework for structured PPE compliance Reporting in clinical scenes

Conventional cameras capture images, but PPE auditing in healthcare requires structured per-person compliance records rather than image-level detections alone. In this work, we present an applied, person-centric framework that bridges instance-level detections to structured per-person PPE records, enabling interpretable, automation-ready outputs. The pipeline uses detector-derived person anchors from head detections, assigns each PPE instance to at most one person using an anchor-gated geometric association rule for clinical scenes with crowding and overlap, and maps the assigned evidence to discrete compliance states that match the dataset encoding, producing a strict JSON report. As an optional extension, we also study schema-constrained projection for cases where a vision–language model (VLM) is used for direct report generation under strict schema requirements. We evaluate on 1908 images with 5791 person instances spanning non-dense and dense scenes. The deterministic prediction-derived reports achieve around 90% agreement with ground truth across core PPE fields, and the safety-critical any_noncompliance decision reaches 90.0% accuracy with 93.6% precision, 90.4% recall, and a 92.0% F1 score. Across the density splits considered here, performance remains strong within the current evaluation set, and for the optional VLM-based reporting extension, schema-constrained projection improves strict schema validity from 96.5% to 99.0%. Overall, the findings support the practical utility of structured, interpretable per-person PPE reporting under the current dataset and evaluation protocol using a simple, transparent, and auditable pipeline, with schema-valid outputs suitable for downstream automation.