An Epipolar Geometry Guided Feature Matching Method for Absolute Visual Localization of UAV

Autonomous localization is a fundamental requirement for uncrewed aerial vehicle (UAV) navigation tasks. In environments where the Global Navigation Satellite System is unavailable, visual localization utilizing onboard cameras emerges as the primary technique. Although vision-based localization methods have been effectively applied, their accuracy is challenged by two primary issues. First, substantial viewpoint differences between UAV images acquired at large pitch angles and nadir-view satellite imagery complicate the matching process. Second, significant illumination changes and seasonal variations, arising from the long update intervals of satellite images, severely impact matching precision. To address these challenges, this article proposes EpiFormer, a detector-free feature matching network guided by epipolar geometry. The method utilizes a local aware transformer encoder to effectively fuse global and local features. Furthermore, it incorporates an epipolar attention mechanism to ensure geometrically consistent matching between oblique UAV imagery and satellite orthoimagery. Finally, a scale-aligned bidirectional adjustment approach is introduced for fine-level refinement of the matching results. Extensive experiments demonstrate that EpiFormer achieves robust and accurate localization, with horizontal localization error not exceeding 10 m under challenging conditions, such as 50^◦ pitch angle, different seasons, and varying lighting, significantly outperforming other deep learning-based methods.