Scholarlink

In response to the difficulties posed by insufficient illumination, complex backdrops, and object obstructions in cable tunnel settings, this paper proposes HT-YOLO—a lightweight model for helmet detection that builds upon a modified You Only Look Once version 10n (YOLOv10n). The primary contributions of this work are summarized as follows. First, a novel Partial Self-Attention Efficient Multi-scale Attention (PSA_EMA) module is developed by substituting the Multi-Head Self-Attention (MHSA) component for an Efficient Multi-scale Attention (EMA) mechanism, thereby significantly improving multi-scale feature fusion. Second, a Multi-dimensional Cooperative Attention (MCA) mechanism is incorporated between the neck and the detection head to augment the representation of spatial-channel features. Furthermore, the Normalized Wasserstein Distance (NWD) loss function is adopted to reduce the positioning sensitivity for small and occluded targets, addressing a key challenge in dense scenes. Finally, GhostNetV2 convolutional modules are deployed to achieve a more lightweight design. Experimental results demonstrate that compared to YOLOv10n baseline, HT-YOLO achieves a 15.8% reduction in parameters, 9.52% decrease in floating-point operations (FLOPs), and 8.7 percentage point improvement in mean Average Precision at IoU threshold 0.5 (mAP@0.5) reaching 90.8%. This framework robustly enhances safety helmet detection accuracy in complex tunnel scenarios while maintaining real-time deployment efficiency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Helmet Wearing Detection Method in Complex Cable Tunnels Based on HT-YOLO Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xianwei Ma,
  • Yang Zhao,
  • Tian Guo,
  • Yingqiang Shang,
  • Kang Xie,
  • Jing Zhang

摘要

In response to the difficulties posed by insufficient illumination, complex backdrops, and object obstructions in cable tunnel settings, this paper proposes HT-YOLO—a lightweight model for helmet detection that builds upon a modified You Only Look Once version 10n (YOLOv10n). The primary contributions of this work are summarized as follows. First, a novel Partial Self-Attention Efficient Multi-scale Attention (PSA_EMA) module is developed by substituting the Multi-Head Self-Attention (MHSA) component for an Efficient Multi-scale Attention (EMA) mechanism, thereby significantly improving multi-scale feature fusion. Second, a Multi-dimensional Cooperative Attention (MCA) mechanism is incorporated between the neck and the detection head to augment the representation of spatial-channel features. Furthermore, the Normalized Wasserstein Distance (NWD) loss function is adopted to reduce the positioning sensitivity for small and occluded targets, addressing a key challenge in dense scenes. Finally, GhostNetV2 convolutional modules are deployed to achieve a more lightweight design. Experimental results demonstrate that compared to YOLOv10n baseline, HT-YOLO achieves a 15.8% reduction in parameters, 9.52% decrease in floating-point operations (FLOPs), and 8.7 percentage point improvement in mean Average Precision at IoU threshold 0.5 (mAP@0.5) reaching 90.8%. This framework robustly enhances safety helmet detection accuracy in complex tunnel scenarios while maintaining real-time deployment efficiency.