Abstract: Real-time detection of mine personnel is an essential part of the construction of intelligent mine. It is of great significance for mine safety production to realize early warning and linkage control of dangerous areas through video monitoring of underground personnel. At present, the visible light image recognition technology needs to be improved for the identification of personnel in the dim environment of underground coal mine. Aiming at the problems of more noise and image blur in the monitoring video caused by uneven illumination and serious coal dust interference in the underground, this paper proposes an improved YOLOv7 mine personnel detection algorithm. Firstly, aiming at the problem of channel isolation caused by direct splicing of ELAN modules, a complex scene detection method based on channel reorganization and feature attention is proposed. Secondly, since the feature fusion results does not focus on the expected target and the model lacks targeted strategies to improve the detection performance of small targets, an ACmix module is added to the neck multi-scale fusion network to take into account both global and local features, which improves the detection ability of the algorithm for small targets. Finally, efficient Intersection over Union (IOU) loss is introduced to improve the convergence speed of the algorithm and reduce the difference between the height and width of the target frame and the prior frame to achieve more accurate positioning. Through the verification of public pedestrian data sets and self-built mine personnel detection data sets, it is shown that the detection accuracy of the proposed algorithm is 3.1% higher than that of the YOLOv7 model, reaching 89.4%; the recall rate is increased by 3.8% to 86.4%, and the speed is increased by 15.8% to 68.8 FPS, meeting the mine personnel real-time detection work requirements.