Ensuring personal protective equipment (PPE) compliance on construction sites is critical for preventing injuries and fatalities, yet manual inspections are slow, error-prone, and lack real-time capability. This study presents an automated PPE compliance monitoring system that detects safety helmets and goggles using YOLOv7 and YOLOv8 object detection models. A dataset of 5,500 images, sourced from public repositories and field captures, was preprocessed and annotated in Roboflow. Both models were trained on Google Colab for 50 epochs (batch size 16, image size 640×640) and evaluated using mean Average Precision (mAP), precision, recall, and inference speed. YOLOv7 achieved mAP50 (0.893), precision (0.902), recall (0.845), and inference speed (13.5 ms), outperforming YOLOv8 with mAP50 (0.866), precision (0.856), recall (0.812), and 14.6 ms. Class-wise, YOLOv7 detected safety helmets (mAP50 0.947) and non-goggles (mAP50 0.916). Deployed on a laptop camera, YOLOv7 accurately monitored compliance in both static images and live video. Results highlight YOLOv7’s superior accuracy–speed balance for real-time, on-site safety enforcement. Future work will expand environmental diversity in datasets and explore newer YOLO variants to enhance robustness under challenging site conditions.
Shubqizam et al. (Fri,) studied this question.