M. Faris Al Hakim, Hadid Ramadhan, Subhan
Ensuring compliance with personal protective equipment (PPE) is essential for reducing workplace accidents and improving safety in industrial environments. Recent advances in deep learning have enabled real-time vision-based monitoring systems; however, many existing approaches rely on single-source datasets and limited PPE categories, which restrict their generalization in real-world scenarios. This study proposes a YOLOv8-based framework for multi-class PPE detection trained on a combined multi-source dataset covering eight PPE-related categories. The proposed method leverages the anchor-free one-stage architecture of YOLOv8 to enable efficient end-to-end training and real-time inference across heterogeneous data conditions. Experimental results show that the model achieves an overall mean Average Precision at an Intersection over Union threshold of 0.5 (mAP@0.5) of 0.766 and mAP@0.5:0.95 of 0.447, demonstrating strong detection performance across multiple PPE categories. High accuracy is observed for PPE items with distinctive visual features, while challenges remain in detecting context-dependent non-compliance cases. Overall, the proposed framework provides a balanced solution between detection accuracy and computational efficiency, while improving generalization across diverse industrial environments. These findings highlight the potential of multi-source training strategies for developing scalable and practical PPE monitoring systems. © 2026 IEEE.
Universitas Negeri Semarang, Department Of Computer Science, Semarang, Indonesia
