L1 Visual Odometry Component

Detailed Description

L1 Visual Odometry handles the high-frequency, sequential tracking of the UAV. It receives the current frame and the previous frame, using SuperPoint for feature extraction and LightGlue for matching.

It estimates the Relative Pose (Translation and Rotation) between frames. Since monocular VO lacks scale, it outputs a "scale-ambiguous" translation vector or utilizes the altitude prior (if passed) to normalize. It is designed to be robust to low overlap but will flag "Tracking Lost" if matches fall below a threshold.

API Methods

`process_frame`

Input: current_frame: FrameObject, prev_frame: FrameObject | None
Output: L1Result
Description:
1. Extracts features (Keypoints, Descriptors) from current_frame (using Model Registry).
2. If prev_frame exists, matches features with prev_frame using LightGlue.
3. Computes Essential/Fundamental Matrix and recovers relative pose (R, t).
4. Returns L1Result: { relative_pose: Matrix4x4, num_matches: int, confidence: float, keypoints: list }.
Test Cases:
- Frame 1 & Frame 2 (Good overlap) -> High match count, valid pose.
- Frame 1 & Frame 10 (Zero overlap) -> Low match count, "Tracking Lost" status.

`reset`

Input: void
Output: void
Description: Clears internal history/state. Used after a "Kidnapped Robot" reset.

Integration Tests

Trajectory Generation: Feed a sequence of 10 frames. Integrate relative poses. Compare shape of trajectory to ground truth (ignoring scale).

Non-functional Tests

Inference Speed: Must run < 100ms (TensorRT) or reasonable time on CPU.

1.7 KiB Raw Blame History