Visual geo-localization and attitude estimation using satellite imagery and topographical elevation for unmanned aerial vehicles

In recent years, the autonomous flight capability of unmanned aerial vehicles (UAVs) has significantly improved. However, basic issues such as geo-localization and orientation still remain unsolved when Global Positioning System (GPS) is unavailable. Current visual geo-localization methods either assume a flat flight area using only satellite imagery for large-scale scenes or rely on three-dimensional (3D) reconstruction for small-scale areas, both of which limit practical UAVs applications. To achieve visual geo-localization in largescale scenes with complex terrain variations, we propose a method that reliably achieves 6-degree-of-freedom pose estimation, including offline database and online inference. In the first stage, neural networks extract global and local features from satellite images to build feature databases. In the second stage, a fast search is conducted in the global feature database based on the global image descriptor of the aerial image, followed by fine matching based on local keypoint descriptors. Geographic coordinates for matched feature points are provided by satellite imagery and topographical elevation to achieve the UAV pose estimation. We conducted experimental validation by capturing aerial images at two non-overlapping locations. The results show that at Location 1, when using a top-down view, the overall localization error of the UAV decreased from 31.76 m to 17.94 m, improving accuracy by 43.61%; the recall rate increased from 67.59% to 71.94%. At Location 2, when using a front-down view, the overall localization error of the UAV decreased from 48.07 m to 10.84 m, with accuracy significantly improving by 77.45%; the recall rate increased from 64.96% to 69.71%.