Abstract: The trackless tyred vehicle is widely used in coal mine auxiliary transportation system because of its good mobility and high efficiency. Facing the dim and complex scene of coal mine, due to the poor lighting conditions and complex scene in coal mine, the accidents of the underground moving equipment like the trackless tyred vehicle occur frequently, seriously threatening the life safety of operators in the underground coal mine. A reliable registration method of infrared and visible dual-modal images is designed and proposed to solve the problem of reliable identification of intruders in front of the moving equipment in underground coal mines. By introducing the Sobel residual block structure, a fusion model of infrared and visible images based on improved STDFusionNet network architecture is constructed to enhance the feature representation of the fused image. Then, a reliable identification method of underground personnel based on YOLOv5s is proposed, and the visible light and infrared dual-modal features are combined to fuse and complement each other, and an end-to-end single-stage detection model of intrusive personnel in front of underground moving equipment is constructed, which realizes the reliable detection of intrusive personnel in front of underground moving equipment. Finally, after the work that the personnel targets are labeled in the actual mining scene of coal mine and the sample expansion is carried out, the personnel sample set along the auxiliary transportation is established. The industrial applicability of the proposed personnel target detection algorithm is verified through industrial experiments. The experimental results show that the improved fusion model can effectively improve the accuracy of object detection compared with the commonly used DenseFuse, FusionGAN and STDFusionNet models. When it is combined with the YOLOv5s object detection architecture, the constructed reliable identification method for intruded personnel in front of underground mobile equipment operation has the accuracy rate of 96.43% and the recall rate of 98.18%, and it also shows obvious advantages in the detection frame rate.