Regulation-aligned PPE compliance assessment for work-at-height using visual relationships and scene graph reasoning

Work-at-height (WAH) poses significant safety challenges on construction sites, where the proper use of personal protective equipment (PPE) is critical for accident prevention. This paper developed an automated framework that integrates visual relationship modeling with scene graph analysis to detect PPE compliance of construction workers. An improved YOLOv11 model with GhostConv and CBAM modules is used to detect small and occluded PPE items (e.g., lifeline and hook). Then, a multi-feature relationship model is developed to infer semantic interactions by fusing semantic, geometric, visual appearance, and global contextual features. These relationships are structured into scene graphs and compared with rule-based graph derived from safety regulations to identify non-compliance. A safety compliance analysis system is further implemented to visualize results and generate structured safety reports. The experimental results demonstrate that the proposed models outperform baselines, achieving a [email protected] of 0.924 and Recall@100 of 0.89. The practical compliance system is developed with visual overlays, structured reports, and actionable insights, offering a scalable and interpretable solution for construction safety supervision.