A research team has developed a robust visual localization framework that lets Unmanned Aerial Vehicles (UAVs) estimate their absolute position without relying on the Global Navigation Satellite System (GNSS). The framework pairs a pre-trained, zero-shot aerial–satellite image matcher—the Generalizable Image Matcher (GIM)—with a particle filter. Front-end visual odometry supplies a continuous position prior, while back-end satellite image matching provides absolute location cues, and an adaptive weighting function fuses the two. Tested across rural, urban, farmland and canyon flights, the approach maintained stable, meter-level positioning without cumulative drift.
Unmanned aerial vehicles (UAVs) have become indispensable in military, agricultural and industrial operations, and their autonomy rests heavily on reliable navigation. Most UAVs today depend on Global Navigation Satellite System (GNSS) for positioning. Yet in urban canyons, conflict zones and forested or mountainous terrain, GNSS signals can be blocked by vegetation or deliberately disrupted by jamming and spoofing—as demonstrated during the 2020 Nagorno-Karabakh conflict. Vision-based localization offers a low-cost alternative, but relative methods such as visual odometry accumulate drift over long flights, while absolute methods that match UAV imagery to satellite maps struggle with differences in viewpoint, resolution, season and illumination. Based on these challenges, in-depth research is needed on a localization framework that fuses relative motion priors with intermittent absolute cross-view observations to maintain both availability and accuracy.
Researchers have proposed a robust visual localization framework that integrates the pretrained GIM_DKM matcher with a particle filter. The study was published (DOI: 10.1186/s43020-026-00232-w) in Satellite Navigation on 02 November 2026. The team adopted the GIM_DKM, a pretrained model derived from DKM using the GIM training framework as the back-end matcher and paired it with particle-filter fusion. The framework was evaluated on flight trajectories spanning rural, urban, farmland and canyon scenarios, including both real-flight and synthetic datasets.
The pipeline has two parts. A lightweight front end uses SuperPoint and LightGlue to match consecutive UAV frames and estimate relative motion, producing a continuous position prior. Guided by this prior, the back end scatters particles on a uniform grid, crops corresponding satellite sub-images, and matches each against the UAV image with GIM_DKM. Random Sample Consensus (RANSAC) removes outlier correspondences before homography estimation and geographic coordinate conversion. A reliability weighting function then scores each candidate by combining its distance from the visual-odometry prior, the number of matched feature points, and the spatial spread of matches, measured as normalized convex-hull area in both images. The weighted average yields the final position. Notably, GIM_DKM was used zero-shot—without fine-tuning on any flight data. Unlike single-frame matching, which fails under seasonal appearance changes or partial cloud and fog occlusion, the particle-based search tolerated such conditions, and local sampling around the prior avoided exhaustive traversal of the reference map.
The authors said the framework was built to keep absolute positioning available when cross-view matching becomes intermittent. They said the key was letting front-end visual odometry constrain where the back end searches, so the matcher only examines a small neighborhood instead of the whole reference map. They said weighting candidate matches by both feature quantity and spatial spread proved more reliable than trusting any single frame-to-satellite match. According to the authors, the experiments showed that even when visual-odometry drift accumulated along some trajectories, the particle distribution progressively concentrated around the true position.
The framework targets outdoor operations in which GNSS is unavailable or unreliable but the ground remains observable and georeferenced satellite imagery exists. With the initial UAV position known—from takeoff or the last reliable GNSS fix—the system sustains meter-level positioning over long flights without drift accumulation, achieving a high localization success rate where single-frame matching often fails. It is relevant to inspection, mapping, agriculture and missions in contested airspace. The authors plan to extend the work to unknown initial positions, night and heavy-weather conditions, and lightweight onboard deployment on resource-constrained UAV platforms.
###
References
DOI
Original Source URL
https://doi.org/10.1186/s43020-026-00232-w
Funding i nformation
National Natural Science Foundation of China (62573232, 62503224); Natural Science Foundation of Jiangsu Province (BK20241388); Chengdu Key R&D Support Program (2025-XT00-00024-GX)
About Satellite Navigation
Satellite Navigation (ISSN: 2662-1363; ISSN: 2662-9291) Satellite Navigation is the official journal of the Aerospace Information Research Institute . The journal aims to report innovative ideas, new results, and progress in the theories, techniques, and applications of satellite navigation. The journal welcomes original articles, reviews and commentaries.
Satellite Navigation
Not applicable
A robust visual localization framework for UAVs using a zero-shot aerial-satellite image matcher and a particle filter
2-Nov-2026
The authors declare that they have no competing interests.