School of Resource Environment and Safety Engineering, University of South China

基于视觉图像方法是目前矿井人员佩戴安全帽智能检测的热点，但现有方法所用的地下矿山数据较少，安全帽特征分类不够精确。通过采集地下矿山采场、井巷等实际生产场景的图像，构建了矿山安全帽佩戴数据集——MHWD，并将安全帽佩戴情况进一步细分为正确佩戴、不规范佩戴和未佩戴3类。采用YOLOX算法进行人员佩戴安全帽检测，为了增强YOLOX提取全局特征的能力，引入注意力机制，即在YOLOX骨干网的空间金字塔池化瓶颈层嵌入有效通道注意力模块，在路径聚合特征金字塔网络每个上采样和下采样后添加卷积块注意力模块，由此构建了YOLOX-A模型。采用MHWD训练YOLOX-A模型并进行验证，结果表明，针对照度低、模糊、有人员遮挡的矿井图像，YOLOX-A模型能够准确识别人员佩戴安全帽情况，对不规范佩戴、正确佩戴和未佩戴安全帽3种分类结果的F1分数分别为0.86,0.92,0.89，平均精度分别为93.16%,95.76%,91.69%，平均精度均值为93.54%，整体F1分数较YOLOX模型提升4%，检测精度高于主流目标检测模型EfficientDet,YOLOv3,YOLOv4,YOLOv5,YOLOX。

Visual image-based methods are currently a hot topic in intelligent detection of mine personnel wearing safety helmets. However, existing methods use limited underground mining data and the classification of safety helmet features is not accurate enough. By collecting images of actual production scenes such as underground mining sites and roadways, a mining helmet wearing dataset (MHWD) is constructed. The helmet wearing situation is further divided into three categories: correct wearing, non-standard wearing, and non wearing. YOLOX algorithm is used to detect personnel wearing helmets. In order to enhance YOLOX's capability to extract global features, the attention mechanism is introduced. The effective channel attention module is embedded in the spatial pyramid pooling bottleneck layer of YOLOX's backbone network. The convolutional block attention module is added after each upsampling and downsampling of the path aggregation feature pyramid network, thus the YOLOX-A model is built. By using MHWD, the YOLOX-A model is trained and validated. The results show that the YOLOX-A model can accurately identify the wearing of safety helmets by personnel in mine images with low illumination, blurriness, and personnel obstruction. The F1 scores for the classification results of non-standard wearing, correct wearing, and non wearing safety helmets are 0.86, 0.92, and 0.89, respectively. The average precision is 93.16%, 95.76%, and 91.69%. The average precision mean is 93.54%. The overall F1 score is 4% higher than the YOLOX model. The detection precision is higher than the mainstream target detection models EfficientDet, YOLOv3, YOLOv4, YOLOv5 and YOLOX.

湖南省研究生科研创新项目（CX20200916,QL20210216,QL20230233）；

杜青,杨仕教,郭钦鹏等.地下矿山作业人员佩戴安全帽智能检测方法[J].工矿自动化,2023,49(07):134-140.DOI:10.13272/j.issn.1671-251x.2022090033.

DU Qing, YANG Shijiao, GUO Qinpeng, et al. Intelligent detection method of working personnel wearing safety helmets in underground mine[J]. Journal of Mine Automation，2023，49(7)：134-140.