Scholarlink

<p>In complex and ever-changing mining environments, frequent safety accidents highlight the growing importance of accurately detecting personal protective equipment (PPE). However, traditional detection methods struggle to adapt to diverse equipment types and complex environmental conditions, often resulting in low accuracy and real-time performance due to human factors such as monitoring fatigue-induced inconsistency, subjectivity in visual inspection, and limited coverage of manual supervision. To address these challenges, this paper proposes a real-time PPE detection method based on YOLOv8, aiming to enhance detection performance in complex environments. By introducing EfficientViT as the backbone network, we improve real-time processing efficiency while maintaining high detection accuracy. The self-correcting illumination network (SCINet) is integrated into the bacbone to adaptively enhance image contrast under low-light conditions and mitigate glare interference. In the neck network, multi-scale dilated attention (MSDA) captures multi-scale features through dilated convolutions, reducing redundancy in global attention mechanisms. To strengthen shape perception for irregular PPE, standard convolutions are replaced with dynamic snake convolution (DSConv). Finally, spatial and channel reconstruction convolution (SCConv) is incorporated into the detection head to suppress feature redundancy, further improving both detection accuracy and real-time performance. Experimental results demonstrate that our proposed model achieves 94.7% accuracy (4.1% improvement) and 85 FPS (23% improvement) with only 2.5 million parameters (22% reduction) on the self-constructed dataset, effectively improving real-time detection of PPE in complex mining environments.</p>

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

Real-time personal protective equipment detection and classification with YOLOv8 multi-scale fusion

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Zheng Wang,
  • Yingjie Zhang,
  • Shilong Zhang

摘要

In complex and ever-changing mining environments, frequent safety accidents highlight the growing importance of accurately detecting personal protective equipment (PPE). However, traditional detection methods struggle to adapt to diverse equipment types and complex environmental conditions, often resulting in low accuracy and real-time performance due to human factors such as monitoring fatigue-induced inconsistency, subjectivity in visual inspection, and limited coverage of manual supervision. To address these challenges, this paper proposes a real-time PPE detection method based on YOLOv8, aiming to enhance detection performance in complex environments. By introducing EfficientViT as the backbone network, we improve real-time processing efficiency while maintaining high detection accuracy. The self-correcting illumination network (SCINet) is integrated into the bacbone to adaptively enhance image contrast under low-light conditions and mitigate glare interference. In the neck network, multi-scale dilated attention (MSDA) captures multi-scale features through dilated convolutions, reducing redundancy in global attention mechanisms. To strengthen shape perception for irregular PPE, standard convolutions are replaced with dynamic snake convolution (DSConv). Finally, spatial and channel reconstruction convolution (SCConv) is incorporated into the detection head to suppress feature redundancy, further improving both detection accuracy and real-time performance. Experimental results demonstrate that our proposed model achieves 94.7% accuracy (4.1% improvement) and 85 FPS (23% improvement) with only 2.5 million parameters (22% reduction) on the self-constructed dataset, effectively improving real-time detection of PPE in complex mining environments.