GeoRVLF: A Robust Drone-Satellite Visual Geo-Localization Framework for Small Unmanned Aerial Vehicle Platforms

Abstract

Drone-satellite geo-localization is a novel technology for autonomous UAV positioning in GNSS-denied environments, but it faces many complex challenges such as cross-scale heterogeneous scenes and real-time operational efficiency in practical applications. This letter proposes a robust visual geolocalization framework for small UAV platforms. First, we performed Cross-scale Place Recognition (CPR) module using a multi-level block retrieval algorithm, which could reduce the runtime and storage usage of feature-based VPR method. In the process we constructed a heterogeneous scene matching network HSM-Net, and designed an accurate Fine Pose Registration (FPR) geo-localization module for 4-DoF global pose estimations. Considering the computility limitation of small UAV platforms, we fused the FPR with VO module via pose graph optimization method, which improved the operating efficiency and stability of the framework. Through dataset evaluations and real-world flight experiments, we verified its geo-localization performance. The results show that GeoRVLF can could accurately and stably estimate the global pose of small UAVs in complex heterogeneous scenes, whose trajectory accuracy is close to low-accuracy GPS and operating frequency is near to classic VO methods.