Towards GNSS-Free UAV Localization via XFeatLightGlue and Semantic Embedding

Accurate and robust geo-localization is essential for autonomous UAV navigation in GNSS-denied environments. Vision-based localization methods achieve positioning by matching UAV images with satellite maps, but face challenges from illumination, scale, and viewpoint differences between aerial and satellite imagery. We propose a vision-based UAV localization framework that combines XFeat, a deep learning-based local feature extractor, with LightGlue, a local feature matcher, and further enhances the method’s accuracy and robustness through semantic context embedding and multi-scale feature representation. Experimental results on public datasets demonstrate that our method significantly outperforms traditional feature matching approaches in both accuracy and robustness, achieving performance comparable to GNSS in long-range and high-altitude UAV localization scenarios.