GNSS-Denied Visual Localization for UAVs using Satellite Imagery
SIU 2026. SuperPoint–LightGlue matching of UAV camera frames to satellite tiles; 10 m MAE / 14 m RMSE (X–Y) at ~500 m vs RTK, ~5 Hz on Jetson Orin Nano.
SIU 2026 — Deep Visual Localization for UAVs in GNSS-Denied Environments Using Satellite Imagery (Engin Öksüz, Barış Yalçın, Tolga Demirdal). Proceedings are submitted to IEEE Xplore; the button above is the Xplore search for this title.
Absolute position from a downward/oblique UAV camera matched to open satellite tiles. No GNSS in the estimator. Relative methods (VO / VIO / SLAM) drift; this pipeline re-anchors to a georeferenced map.
Pipeline
Four stages, as in the paper:
- Map synthesis — previous pose (P) and camera (K) select satellite tiles; a homography (H) warps the map toward the camera (altitude, roll, pitch, yaw).
- Matching — CLAHE, then SuperPoint + LightGlue. SIFT and ORB are baselines on the same backend, not alternate front-ends in the deployed system.
- Pose — EPnP + RANSAC on 2D–3D correspondences. Typical operating point: SuperPoint 800–1200 keypoints, LightGlue 80–200 inliers, EPnP uses 15–40.
- Fusion — linear Kalman filter. When matches collapse (low texture), Lucas–Kanade optical flow + Procrustes is the backup. A rate limiter bounds jump updates.
Deployed with TensorRT on a Jetson Orin Nano at about 5 Hz (~200 ms end-to-end). That is the number in the paper — not 10–30 Hz, not AGX.
Field results (RTK truth)
Rural, low-texture flights: about 20 sorties / 12 hours, 50–900 m AGL. Headline accuracy is the ~500 m band, X–Y only:
| MAE | RMSE | |
|---|---|---|
| Proposed (SP + LightGlue) | 10 m | 14 m |
Ablation on a representative segment (X–Y MAE vs RTK):
| Configuration | MAE (m) |
|---|---|
| Full system | 10.2 |
| Homography off | 23.1 |
| Optical-flow backup off | 12.6 |
Matching on the same protocol:
| Front-end | mAP | Time (s) | Matches |
|---|---|---|---|
| SIFT | 0.650 | 0.043 | 24 |
| ORB | 0.717 | 0.036 | 78 |
| SuperPoint + LightGlue | 0.912 | 0.340 | 249 |
North and East tracks follow RTK. Z is worse — DEM resolution and monocular depth, not a solved axis. Homogeneous farmland still drops keypoints and produces short error spikes; Kalman + optical flow damp them, they do not disappear.
What this is not
- Not 4.8 m mean error. That figure was on an older draft of this page and is not in the paper.
- Not a claim of consumer-GPS parity, indoor flight, urban air mobility, or night-as-primary.
- Not a full 12-state EKF with camera intrinsics in the state. The published filter is a linear Kalman on the pose stream.
- Employer name and platform internals are not discussed here.
Figures
SuperPoint–LightGlue correspondences between the camera frame and the warped satellite tile.
Estimated track on the satellite map vs RTK. Paper headline: 10 m MAE / 14 m RMSE in X–Y near 500 m.
NED traces. Horizontal axes track RTK more closely than height.
Next (paper)
Other terrain and seasons, better Z, and a LoFTR (or similar) comparison. Not on this page until measured.