Multi-source geospatial fusion for enhanced visual geo-localization of unmanned aerial vehicles

Currently, although significant advancements have been made in the autonomous navigation of Unmanned Aerial Vehicles (UAVs), reliable geo-localization remains a challenge when satellite navigation signals fail. Existing research primarily focuses on cross-view image retrieval, feature point matching, and visual localization based on three-dimensional (3D) reconstruction. While these methods work effectively in small-scale flight scenarios with pre-acquired image data, they struggle in larger-scale environments lacking prior information. To bridge this gap, we propose a visual geo-localization framework using multi-source geospatial fusion, which consists of two main phases: an offline phase for constructing multi-source databases and an online hierarchical geo-localization phase. During the offline phase, satellite imagery undergoes multi-perspective augmentation through rotational and translational transformations, followed by extraction of global and local features using neural networks to build feature databases. Simultaneously, terrain elevation data within the satellite imagery range is preserved for localization inference during the online phase. The online phase employs cascade matching: first, a coarse search based on global descriptors rapidly narrows potential locations; second, local feature matching obtains the pixel coordinates of matching feature points in aerial images and satellite images; third, geographic coordinates for the matched feature points are derived from satellite imagery and terrain elevation data, and the UAV geo-localization is achieved by leveraging stereoscopic projection geometry. We conducted experimental validation across geographically distinct test sites, covering an area of 26.9 km by 7.03 km, using aerial image captured at approximately 500 m altitude from both nadir view (90^∘) and oblique view (45^∘ tilt). The results confirm the effectiveness of our proposed approach.