Scholarlink

<p>As mining advances, underground safety has become increasingly important. Quickly detecting worker actions helps prevent accidents. To address poor lighting and high dust in coal mines that reduce model accuracy, and the trade-off between accuracy and complexity in behavior recognition models, we propose a lightweight Channel-Position Spatio-Temporal Graph Convolutional Network (Lite-CPAM-STGCN++). Our approach first uses YOLOF, a lightweight object detection algorithm, to detect personnel in coal mines. Then, it combines YOLOF with a lightweight High-Resolution Network (Lite-HRNet) for pose estimation, which extracts skeletal data of underground workers. The use of skeletal data effectively mitigates the impact of complex underground environments on recognition accuracy. Second, leveraging the relatively singular and fixed characteristics of underground coal mine scenarios, we design a lightweight spatio-temporal GCN. By pruning the ST-GCN + + structure and adjusting the number of channels to adapt to the underground environment, the model achieves lightweight properties. Finally, to enhance feature representation capability and focus on key feature information while reducing the model’s learning of irrelevant features, we integrate a Channel-Position Attention Module for GCN (CPAM-GCN). The CPAM attention mechanism fuses channel and position information output by GCN and Temporal Convolutional Networks (TCN). Experimental results on a self-constructed dataset of underground coal mine behaviors showed that the proposed method achieved a Top-1 accuracy of 90.62% for eight types of underground behaviors, with the parameter count reduced to 25% of the original model and an FPS of 75.49, demonstrating good recognition accuracy and speed. Furthermore, to further improve the model’s accuracy, we optimized the fusion strategy of information flow, enhancing the model’s feature extraction capability for skeletal data and achieving a Top-1 accuracy of 91.35%. This method can accurately identify dangerous behaviors of underground workers, providing a scientific basis for reducing safety accidents.</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

The method for underground personnel behavior recognition based on multi-information flow collaborative graph convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Guoxin Wang,
  • Dan Yuan,
  • Hongfang Ru

摘要

As mining advances, underground safety has become increasingly important. Quickly detecting worker actions helps prevent accidents. To address poor lighting and high dust in coal mines that reduce model accuracy, and the trade-off between accuracy and complexity in behavior recognition models, we propose a lightweight Channel-Position Spatio-Temporal Graph Convolutional Network (Lite-CPAM-STGCN++). Our approach first uses YOLOF, a lightweight object detection algorithm, to detect personnel in coal mines. Then, it combines YOLOF with a lightweight High-Resolution Network (Lite-HRNet) for pose estimation, which extracts skeletal data of underground workers. The use of skeletal data effectively mitigates the impact of complex underground environments on recognition accuracy. Second, leveraging the relatively singular and fixed characteristics of underground coal mine scenarios, we design a lightweight spatio-temporal GCN. By pruning the ST-GCN + + structure and adjusting the number of channels to adapt to the underground environment, the model achieves lightweight properties. Finally, to enhance feature representation capability and focus on key feature information while reducing the model’s learning of irrelevant features, we integrate a Channel-Position Attention Module for GCN (CPAM-GCN). The CPAM attention mechanism fuses channel and position information output by GCN and Temporal Convolutional Networks (TCN). Experimental results on a self-constructed dataset of underground coal mine behaviors showed that the proposed method achieved a Top-1 accuracy of 90.62% for eight types of underground behaviors, with the parameter count reduced to 25% of the original model and an FPS of 75.49, demonstrating good recognition accuracy and speed. Furthermore, to further improve the model’s accuracy, we optimized the fusion strategy of information flow, enhancing the model’s feature extraction capability for skeletal data and achieving a Top-1 accuracy of 91.35%. This method can accurately identify dangerous behaviors of underground workers, providing a scientific basis for reducing safety accidents.