Abstract: The underground environment of coal mines is complex and has more safety hazards. Personnel detection is an important part of ensuring safe production in coal mines and building smart mines. Commonly used detection algorithms have large parameter counts, high requirements on equipment arithmetic, and are not satisfactory for application in low illumination environments in coal mines. To address the above problems, a lightweight personnel detection method YOLOv5-CWG is proposed for underground coal mine based on YOLOv5.Firstly, the coordinate attention mechanism (Coordinate Attention) embedded in the backbone network adaptively adjusts the weights of each channel in the feature map to enhance the expression ability of the features. It improves the attention to the personnel target to be detected under the unfavorable conditions of low illumination, serious dust influence and low contrast, and locates and identifies the personnel target more accurately. Secondly, the weighted multiscale feature fusion module (Weighted multiscale feature fusion moule) introduces learnable weights to give different attention to the feature layer. This enables the network to effectively fuse shallow positional features and high-level semantic information to enhance the information extraction capability and better distinguish between target and interference, improving the anti-interference capability of the model. Add a P2 layer detection head to improve the detection and localization accuracy of smaller targets. Introduce SIoU instead of the original loss function to accelerate model convergence. Finally, the introduction of the Ghost module to optimize the backbone network can reduce the computational and parametric quantities of the model without losing the model performance, improve the detection speed, and make the model easier to be deployed on resource-constrained devices. Experimental results show that the proposed YOLOv5-CWG algorithm achieves 97.5% mAP on the Underground Mine Personnel Detection Dataset (UMPDD). Compared with YOLOv5s, the mAP is improved by 7.3%, the computation amount is reduced by 27.6%, and the FPS is improved by 6.3. The proposed algorithm significantly improves the accuracy of personnel detection in underground coal mines, and efficiently solves the problem of difficult personnel detection caused by low brightness and uneven illumination.