Global Map Optimization based Real-time UAV Geolocation without GNSS

Maintaining real-time and stable geolocation for unmanned aerial vehicles (UAVs) using pre-existing geolocation information remains a pivotal and pressing challenge in the realm of autonomous navigation and localization. To tackle this, we introduce a framework that integrates visual Simultaneous Localization and Mapping (vSLAM) with a deep-learning-based UAV-to-satellite (U2S) image matching technique, specifically tailored for downward-tilted camera setups. This framework incorporates a similarity-based heterogeneous image matching approach, enabling 6 degrees of freedom (DOF) U2S prediction for UAVs. Utilizing this prediction, we devise a single-shot global initialization method for rapid and robust global initialization. Furthermore, to bolster localization accuracy, we propose an iterative map fusion method that integrates U2S predictions with visual data from vSLAM in real-time. Evaluations conducted using aerial datasets demonstrate the proposed method’s capability to facilitate rapid and effective global initialization, as well as real-time and accurate estimation of the UAV’s geographical pose.