To address the problems of poor edge segmentation performance and low segmentation accuracy of existing real-time segmentation methods for underground coal mine rail regions under low-light conditions, a high-precision real-time segmentation network (HPRTSNet) was proposed. In the shallow feature extraction stage, a Sobel-based multi-scale edge enhancement module (SMEEM) was designed, in which Sobel operators and multi-scale adaptive average pooling were employed to capture rail edge features, thereby enhancing edge representation of rail regions. In the deep feature extraction stage, a dual-domain fusion enhancement module (DFEM) was introduced to fuse spatial and frequency-domain information, effectively improving the understanding of the complex underground backgrounds. In addition, a receptive-field hybrid atrous spatial pyramid pooling (RHASPP) module was constructed to expand the receptive field and enhance rail feature representation. To further optimize model performance, a hybrid segmentation loss function was adopted. Experimental results on a self-built mine-track segmentation dataset demonstrated that, compared with existing real-time segmentation algorithms, HPRTSNet achieved superior performance in both segmentation accuracy and real-time efficiency.