In order to solve the problem that the specific post-processing steps of the existing downhole personnel detection algorithms, such as Anchor and Non-Maximum Suppression, lead to the complex training process, a detection method based on improved Transformer was proposed to improve the detection accuracy of underground personnel. In this algorithm, Detection Transformer was used as the basic detection framework, and the lightweight Swin Transformer network was used to replace the backbone network ResNet. At the same time, a learnable detection block was added to the input sequence, for which a reconfigured attention module was introduced. In order to integrate multi-scale features, a lightweight, non-encoding neck structure was adopted to reduce the computational costs. Surveillance videos of different working scenarios were collected to create datasets for experiments. The results show that the proposed method has achieved excellent performance on the self-built underground personnel detection dataset, with an average detection accuracy of 97. 8% , which is better than that of other underground personnel detection networks, and the detection speed reaches 32 frames per second.