A Coarse-to-Fine Localization Algorithm for UAVs Based on Heterogeneous Image Matching

To enable UAVs to achieve rapid and precise localization in GPS-denied environments, this paper proposed a coarse-to-fine autonomous localization framework for UAV from the perspective of heterogeneous visual matching. Initially, a QATM-SqueezeNet was devised with the aim of accomplishing coarse localization by template matching between the images captured by UAVs and reference satellite images. Subsequently, a more refined UAV localization algorithm was designed. Based on the approximate position acquired through template matching, accurate localization was achieved by means of graph neural network feature matching. Furthermore, in order to mitigate the impact of non-coplanarity conditions on the localization results, geo-semantic information was introduced to facilitate the filtering of feature points. The efficacy of the algorithm was demonstrated through experimentation conducted in actual typical navigation scenarios. Experimental results indicate that, in visually feature-rich areas, the average localization error was less than 2 m, whereas in visually sparse areas, the error remained within 13 m. It is notable that our approach does not require prior position of UAVs, which offers a promising avenue for UAV visual autonomous localization.