Performance Evaluation of YOLOv8 and YOLOv9 for Safety Helmet Detection in Construction Environments

Ensuring worker safety through compliance with Personal Protective Equipment (PPE) mandates, particularly the use of safety helmets, is a critical concern in the construction industry. Automated surveillance systems leveraging advanced object detection models offer a promising solution to enhance monitoring and reduce accidents. This paper presents a comparative performance evaluation of two promi- nent YOLO (You Only Look Once) variants, YOLOv8 and YOLOv9, for the specific task of safety helmet detection. Utilizing a publicly available hard-hat dataset featuring diverse construction site scenarios, both models were trained for 7 epochs and rigorously evaluated. Performance metrics, including precision, recall, mean Average Precision (mAP), F1-score, and confusion matrices, were analyzed. The experimental results indicate that YOLOv9 exhibits a marginal but consistent performance advantage over YOLOv8, achieving an mAP@0.5 of 66.70% compared to YOLOv8’s 65.73%, and an F1-score of 76.15% versus 75.66%. This study underscores the incremental improvements in the YOLO architecture and provides valuable insights for selecting robust models for real-world safety monitoring applications. While both models demonstrate high precision, the relatively lower recall suggests areas for future improvement through more extensive training or model fine-tuning to enhance detection rates in safety-critical environments.