Abstract: The accurate detection of key equipment and personnel in the fully mechanized working face is an important link to realize the information perception of intelligent coal mining. The traditional target detection algorithm realizes the target detection by extracting the features manually. But it is easily affected by the environment and it is not universal. The target detection algorithm based on the convolutional neural network can extract deep information adaptively. But the detection precision is not high, the network parameters are too many, and the calculation is too large in complex environment. In order to the above problems, an improved YOLOv4 model is proposed and applied to the target detection of the fully mechanized working face. In order to accurately detect targets in the complex environment of a fully mechanized working face, a residual self-attention module is integrated into the CSPDarkNet53 network. The capability of acquiring global information is enhanced while parameter sharing and efficient local information aggregation are ensured. The capability of expressing the features of key targets in an image is improved, and the target detection precision is further improved. In order to meet the requirement of high efficiency of target detection in the fully mechanized working face, depthwise-separable convolution is introduced to replace traditional convolution. The model parameter quantity and calculation quantity are reduced. It is beneficial to the industrial deployment of the model. And it improves target detection speed. The experimental results show that compared with YOLOv3, CenterNet and YOLOv4 models, the average precision of the improved YOLOv4 model is the highest, up to 92.59%. It has better balance in parameter quantity, calculation quantity and detection precision. It can accurately detect the target in the complex environment such as coal dust interference, uneven lighting and motion blur.