Utilizing YOLOv11 for Real-Time Construction PPE Compliance Detection

Personal Protective Equipment (PPE) adherence monitoring remains a persistent challenge in the construction sector. Human oversight has proven ineffective in complex or expansive settings, underscoring the importance of intelligent automation. This study aims to present an object-detection system for real-time PPE monitoring utilizing YOLOv11, the latest version of the You Only Look Once (YOLO) model series. The system categorized compliance and non-compliance into 15 classes, including cover classes and their counterparts. The model was trained on 717 annotated images and evaluated using standard object detection benchmarks, including precision, recall, and mean Average Precision (mAP) at an Intersection over Union (IoU) threshold of 0.5 ([email protected]). Several data augmentation strategies were used during training to strengthen generalization. During the process, the total precision and [email protected] were 0.810, with the strongest performance in detecting the mask and hardhat categories. Additional testing was conducted on visually degraded images, including blurred, grayscale, and Out-Of-Distribution (OOD) areas, to further assess robustness. Although strong performance was maintained under standard conditions, decreased confidence was acquired when detecting smaller, ambiguous objects. Overall, YOLOv11 performed well for real-time PPE detection under dynamic site conditions, but factors such as class imbalance and susceptibility to visual noise indicated the need for further refinement. Optimizing powerful yet lightweight model variants tailored for edge deployment, as well as sophisticated augmentation strategies and robust domain-adaptation frameworks, should be prioritized in future research to improve real-world applicability.