YOLOv10-based multi-scale variant object detection for multi-category PPE non-compliance monitoring on construction sites

초록

Multi-category Personal Protective Equipment (PPE) non-compliance monitoring on construction sites remains challenging because visual cues for missing PPE often appear at multiple scales, particularly as small objects. To address these challenges, an enhanced YOLOv10-based detector is developed that incorporates Omni-Dimensional Dynamic Convolution (ODConv) to strengthen fine-grained feature representation, an Efficient Multi-scale Attention-guided Bidirectional Feature Pyramid Network (EMA-BiFPN) to improve cross-scale fusion, a four-scale prediction strategy to increase sensitivity to extremely small instances, and the Minimum Points Distance IoU (MPDIoU) loss to stabilize bounding-box regression. Experiments are conducted on a dataset of 4060 images annotated for five violation categories (No-helmet, No-mask, No-vest, No-glove, and No-shoes), using stratified splits and a controlled hyperparameter search (243 configurations). On the test set, the proposed method achieves 92.49 mAP@50 and 70.04 mAP@50:95 while maintaining 57.28 FPS, outperforming representative two-stage, one-stage, and transformer-based baselines. The largest gains are observed in small-object regimes, reaching = 82.74 and = 88.96, indicating improved robustness to distant and tiny PPE cues. Robustness analyses further show that performance is less sensitive to weight initialization than to dataset partitioning, and condition-based evaluation confirms that far-field views and worker overlap remain the dominant failure cases due to the difficulty of localizing extremely small instances.

키워드