<![CDATA[Computer vision-based Personal Protective Equipment (PPE) detection has become increasingly important for improving Occupational Safety and Health (OSH) compliance through automated, real-time monitoring of workers' safety practices. Although recent studies have reported promising performance for YOLO-based object detection models, most evaluations have been conducted under standard imaging conditions, providing limited evidence of model robustness against photometric variations. To address this gap, this study proposes a systematic robustness evaluation framework based on controlled photometric saturation variations to compare the performance stability of three state-of-the-art object detection models: YOLOv8, YOLOv11, and YOLOv12, for multi-class PPE detection. The experimental dataset comprised 3,116 images annotated into eight PPE-related classes: Helmet On, No Helmet, Vest On, No Vest, Gloves On, No Gloves, Boots On, and No Boots. To ensure a fair comparison, all models were trained using identical experimental settings and evaluated using Precision, Recall, mAP@50, and mAP@50–95. Model robustness was assessed under three saturation conditions (−30%, 0%, and +30%), representing realistic color variations commonly encountered in construction-site surveillance. The experimental results revealed that photometric saturation variations produced only marginal changes in detection performance across all evaluated models. Among the three architectures, YOLOv8 achieved the highest overall performance, attaining an mAP@50 of 55.1%, compared with 52.9% for YOLOv11 and 48.6% for YOLOv12, while maintaining the most stable performance under varying saturation levels. Although YOLOv12 demonstrated relatively better capability for detecting several small-object classes, its overall detection performance remained inferior to that of YOLOv8. These findings indicate that YOLOv8 provides the best trade-off between detection accuracy and robustness under moderate photometric saturation variations. This study contributes a systematic robustness evaluation of recent YOLO architectures under controlled photometric conditions and offers practical insights for selecting reliable object detection models for real-world PPE monitoring systems.]]>
