The intrusion of foreign objects on underground railways in coal mines seriously affects transportation safety. Aiming at the poor effect and low real-time performance of the current algorithms on the segmentation of foreign body edge information, an asymmetric codec-structure method for the segmentation of underground track foreign body is proposed. Firstly, an efficient multi-scale feature extraction backbone network is designed based on Transformer structure in the coding stage to improve computing efficiency while maintaining accuracy. Secondly, the shallow feature enhancement module is proposed to solve the problem that it is difficult to accurately partition the boundary of foreign bodies in the mine. After that, the lightweight Concat decoder is used to fuse the feature information of different scales and predict the segmentation results. Finally, the mixed loss is designed to improve the sensitivity and accuracy of the network to various foreign bodies in the underground environment. The experimental results show that the average interaction ratio, average pixel accuracy and segmentation rate of the proposed method are all the best, which are 86.83%, 92.49% and 36.9fps, which meet the requirements of high accuracy and real-time in the task of underground track foreign body segmentation.