The detection of underground personnel is one of the key technologies in computer vision. However, this detection technique is susceptible to complex environments, resulting in low accuracy and slow speed. To accurately detect underground coal mine operators in complex environments, we combine the underground image features with K-means++ clustering anchor frames and propose a new Re-parameterization YOLO (Rep-YOLO) detection algorithm. First, the Criss-Cross-Vertical with Channel Attention (CVCA) mechanism is introduced at the end of the network to capture the Long-Range Dependencies (LRDs) in the image. This mechanism also emphasizes the significance of different channels to enhance image processing performance and improve the representation ability of the model. Second, the new Deep Extraction of Re-parameterization (DER) backbone network is designed, which adopts the re-parameterization structure to reduce the number of parameters and computation of the model. Additionally, each DER-block fuses different scales of features to enhance the accuracy of the model’s detection capabilities. Finally, Rep-YOLO is optimized using a slim-neck structure, which reduces the complexity of the Rep-YOLO while maintaining sufficient accuracy. The results showed that the Rep-YOLO model proposed in this paper achieved an accuracy of \(87.5\%\), a recall rate of \(77.2\%\), an Average Precision (AP) of \(83.1\%\), and a Frame Per Second (FPS) of 71.9. Compared to eight different models, the recall, AP50, and FPS of the Rep-YOLO model were improved. The research shows that the Rep-YOLO model can provide a real-time and efficient method for mine personnel detection. Source code is released in https://github.com/DrLSB/Rep-YOLO.