Comparison of YOLO algorithms for PPE compliance monitoring at construction sites

Abstract

Safety is a critical concern in the construction industry, where workers are exposed to various hazards that can lead to serious injuries or fatalities. Personal Protective Equipment (PPE) is vital in protecting the workers and increasing their safety. However, ensuring consistent PPE compliance among construction workers remains a challenge. To overcome this challenge, this study developed automated PPE compliance monitoring models through the You Only Look Once (YOLO) object detection algorithm. The variants of YOLO algorithms such as YOLOv8-s, YOLOv9-s, and YOLOv11-s were trained to identify the best performance model to detect and classify the presence of humans and four major PPE items: helmets, high-visibility vests, gloves, and boots. To prevent the overfitting of the models, early stopping with a patience level of 20 epochs was set to the models. The system was externally validated using the construction sites’ images to check its applicability to the Sri Lankan context. The efficacy of each model was assessed using performance evaluation matrices such as precision-recall curves, mean Average Precision (mAP@0.5), inference time, and loss function values. The results show that YOLOv9-s outperformed the other models in overall performance even though, it went through the highest number of epochs during training. Future work can explore enhancing the YOLOv9-s model performance by integrating motion detection through IoT devices, allowing for more precise tracking of PPE compliance and reducing false detections of idle or stored PPE. This approach could significantly improve real-time monitoring of PPE compliance for worker safety at construction sites.