gps-denied-onboard/docs/02_components/07_L3_Metric_Refinement/spec.md

L3 Metric Refinement Component

Detailed Description

L3 Metric Refinement provides the "Pinpoint" accuracy. It takes a pair of images: the current UAV frame and a candidate Satellite Tile (provided by L2 or determined by the State Estimator's prediction).

It uses LiteSAM to find dense correspondences between the aerial and satellite views. It computes a Homography and solves for the absolute world coordinate of the UAV center, achieving pixel-level accuracy.

API Methods

refine_pose

• Input: uav_frame: FrameObject, satellite_tile: SatelliteTile, prior_yaw: float (optional)
• Output: L3Result
• Description:
  1. Preprocesses images for LiteSAM.
  2. Infers dense matches.
  3. Filters outliers (RANSAC).
  4. Computes Homography H.
  5. Decomposes H into rotation and translation (using camera intrinsics and tile GSD).
  6. Returns { abs_lat: float, abs_lon: float, yaw: float, confidence: float, match_inliers: int }.
• Test Cases:
  • Perfect match pair -> Accurate Lat/Lon.
  • Mismatched pair (wrong tile) -> Low inlier count, Low confidence.

Integration Tests

• Accuracy Verification: Use a ground-truth pair (UAV image + known correct Satellite tile). Verify calculated Lat/Lon error is < 20 meters.

Non-functional Tests

• Robustness: Test with satellite images from different seasons (if available) or synthetically altered (color shift) to verify LiteSAM's semantic robustness.