qAicedrone-Roads. A Robust Tool for Road Marking Extraction Using Aerial Photogrammetry and U-Net

Abstract

Efficient and accurate road marking detection is essential for infrastructure maintenance, traffic management, and the development of digital twins for autonomous mobility. However, most existing methods rely on orthomosaics or single-image detections, which suffer from geometric distortions and occlusions and have limited semantic insight. To address these limitations, this study introduces qAicedrone-Roads, an open-source tool integrated into QGIS that enables the automatic detection, classification, and mapping of road markings from UAV-based photogrammetric imagery. The methodology combines U-Net-based semantic segmentation, a multiview photogrammetric approach, and alignment with a national road marking catalog to enhance geometric accuracy and assign semantic labels. Applied to a real-world case study, the tool achieved high precision, with F1-scores of 0.92 for nonlinear and 0.93 for linear markings, outperforming traditional single-view Computer Vision (CV) methods. These results demonstrate the tool's robustness and accuracy in complex urban environments, enabling the efficient generation of detailed road marking datasets. By facilitating the scalable and reproducible creation of digital twins, qAicedrone-Roads supports smart infrastructure monitoring and sustainable urban mobility planning.