A Direction-Aware Lightweight Network for Camera-Based Underground Mine Track Region Segmentation

Accurate localization of the visible track region is essential for perception using front-mounted cameras on underground rail-guided mine vehicles. The task is difficult because the track foreground occupies only a small image area, and its boundary appearance changes with illumination, water, dust, and scene clutter. This study formulates local perception of the track corridor as binary semantic segmentation of the surface bounded by the two visible rails. RailDLA is a lightweight encoder–decoder network. It combines track context preconditioning, RDLA directional strip propagation, context-guided feature fusion, and track axis proxy decoding. On a self-constructed dataset of underground mine vehicle imagery, RailDLA achieves 96.50% mIoU, 92.10% track IoU, 97.50% track accuracy, and 99.70% pixel accuracy. On the working split, its track IoU exceeds those of FastSCNN, PIDNet-S, DDRNet-23-slim, and SegNeXt-S by absolute margins of 6.06, 1.70, 1.56, and 0.37 percentage points, respectively. Under the unified runtime protocol, RailDLA reaches 120.00 FPS on an NVIDIA GeForce RTX 3070 Laptop GPU. These results demonstrate accurate, real-time inference for underground mine vehicle perception.