ViGSeg: A Vision-GNN Enhanced Visible-light Airport Lane and Pavement Segmentation Network

Semantic segmentation of airport surface elements, such as lane markings and pavement, is essential for accurate perception in aviation environments. In this paper, we propose ViGSeg, a graph-enhanced semantic segmentation framework that incorporates multiscale dilated feature aggregation (MDFA) and graph-guided structural loss to improve region consistency and structural accuracy. To support this task, we construct the Airport-Surface Dataset, which includes 800 high-resolution visible-light images captured from a simulated cockpit perspective, annotated at the pixel level for lane markings and drivable areas. Experimental results demonstrate that ViGSeg achieves strong performance in both drivable area and lane segmentation tasks. Specifically, on the BDD100K and our Airport-Surface dataset, ViGSeg achieves an mIoU of 88.97% and 96.46% for drivable area segmentation, respectively, and an IoU of 39.38% and 79.17% for lane segmentation.