Scholarlink

<p>The real-time detection of personnel is a significant component of the construction of intelligent mines, especially for the personnel safety pre-warning in underground excavation workface. The underground excavation environment faces many critical challenges, including noise interference, uneven illumination, and the occlusion of mechanical equipment. Due to traditional&#xa0;detection algorithms' low detection accuracy and significant resource consumption, this study proposes an accurate and lightweight RAD-YOLOv8s (You Only Look Once v8) personnel detection algorithm for personnel safety pre-warning in underground excavation workface. First, a parameterized backbone network based on HGNetv2 (Hierarchical Graph Network) is employed to reduce the algorithm’s complexity while collecting richer feature information. Second, according to the different targets, the C2f-AKConv (Alterable Kernel Convolution) module is introduced to the algorithm’s neck network to flexibly adjust the size and shape of the convolution kernel, enhancing the algorithm's capacity to adapt to target deformation. Finally, a novel DCNV4-Dyhead (Deformable Convolutional Network v4 -Dynamic Head)module is developed to compensate for the limitations of classic standard convolution in long-range modeling and adaptive spatial aggregation to improve model detection performance. The proposed RAD-YOLOv8s detection algorithm was verified on a dataset of underground excavation workface people. The results showed that the mAP@0.5 and GFlops achieved 91% and 21.1%, which were 2.1% and 25.7% higher than YOLOv8s, respectively. Thus, the algorithm demonstrates improved accuracy and efficiency for real-time personnel safety monitoring in smart mines.</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

Improved RAD-YOLOv8s deep learning algorithm for personnel detection in deep mining workings of mines

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Yan song Zhu,
  • Xing Wang,
  • Chao hong Tang,
  • Cun yu Liu

摘要

The real-time detection of personnel is a significant component of the construction of intelligent mines, especially for the personnel safety pre-warning in underground excavation workface. The underground excavation environment faces many critical challenges, including noise interference, uneven illumination, and the occlusion of mechanical equipment. Due to traditional detection algorithms' low detection accuracy and significant resource consumption, this study proposes an accurate and lightweight RAD-YOLOv8s (You Only Look Once v8) personnel detection algorithm for personnel safety pre-warning in underground excavation workface. First, a parameterized backbone network based on HGNetv2 (Hierarchical Graph Network) is employed to reduce the algorithm’s complexity while collecting richer feature information. Second, according to the different targets, the C2f-AKConv (Alterable Kernel Convolution) module is introduced to the algorithm’s neck network to flexibly adjust the size and shape of the convolution kernel, enhancing the algorithm's capacity to adapt to target deformation. Finally, a novel DCNV4-Dyhead (Deformable Convolutional Network v4 -Dynamic Head)module is developed to compensate for the limitations of classic standard convolution in long-range modeling and adaptive spatial aggregation to improve model detection performance. The proposed RAD-YOLOv8s detection algorithm was verified on a dataset of underground excavation workface people. The results showed that the mAP@0.5 and GFlops achieved 91% and 21.1%, which were 2.1% and 25.7% higher than YOLOv8s, respectively. Thus, the algorithm demonstrates improved accuracy and efficiency for real-time personnel safety monitoring in smart mines.