Vision-based lane detection using UAV aerial imagery

This study addresses drone-view lane detection challenges: slender narrow lanes, cluttered urban backgrounds, and extreme foreground-background imbalance. We propose a method for digitizing road network structures based on an improved U-Net model and data post-processing. The method enhances recognition performance by incorporating Convolutional Block Attention Module (CBAM) to suppress irrelevant information, introducing Coordinate Convolution for improved coordinate sensitivity, and adopting a joint loss function based on Top-K selection to alleviate sample imbalance. The optimized model achieves 84.15% Precision, 86.23% MPA and 75.66% MIoU on urban scenarios. Post-processing steps including direction-adaptive elliptical DBSCAN clustering, facilitate the digitization of road network structures, providing vital support for high-definition (HD) map and intelligent urban transportation system management.