A robust visual localization framework for UAVs using a zero-shot aerial-satellite image matcher and a particle filter

Positioning of Unmanned Aerial Vehicles (UAVs) normally relies on the Global Navigation Satellite System (GNSS). However, GNSS signals are prone to blockage, interference, and spoofing, which can cause positioning failures and severely degrade mission performance. Aerial-satellite image matching using intelligent methods (e.g., deep neural networks) has emerged as a promising approach for achieving meter-level absolute positioning accuracy in GNSS-denied environments. To advance localization based on aerial-satellite image matching, we propose a robust localization framework that integrates a zero-shot aerial-satellite image matcher, termed the Generalizable Image Matcher (GIM), with a particle filter. First, the front-end visual odometry estimates the UAV trajectory from consecutive onboard images to produce an initial pose. Next, guided by the front-end estimate, candidate locations are sampled to crop the corresponding satellite sub-images. GIM is then used to match aerial and satellite images and infer absolute location cues. Finally, an adaptive weighting function is designed to fuse the front-end and back-end localization information, thereby yielding the final positioning result. Experiments were conducted using data collected along 12 flight trajectories spanning rural, urban, and farmland scenarios. Assuming that the initial position of each trajectory is known, the proposed method achieved an overall average positioning error of 5.02 m across all 12 trajectories, with average errors of 5.46 m for the nine real-data trajectories and 3.69 m for the three synthetic-data trajectories. Across the evaluated trajectories, which covered a cumulative flight distance of 104,566 m, no accumulation of positioning drift was observed. These results indicate stable positioning performance under the outdoor datasets and experimental conditions evaluated in this study.