# Camera and grip-bar correction audit

The original analysis was wrong to use one fixed horizontal line at `y=607 px` and no camera-motion compensation. The visible active grip bar is perspective-sloped and moves substantially in image coordinates as the handheld camera pans/tilts.

## Measured correction

- Static-background SIFT + RANSAC camera transform at 10 Hz, athlete masked.
- Camera x translation: -152.3 to 11.8 px.
- Camera y translation: -41.0 to 166.6 px.
- Camera rotation: -0.57 to 1.44 degrees.
- Camera scale: 0.9813 to 1.0143.
- Tracked active-bar center y: 481.5 to 696.0 px.
- Tracked active-bar slope: 0.1356 to 0.1846.
- Direct paired-edge bar detections: 153/258 sampled frames.
- Direct-track residual: 1.44 px RMSE, 2.67 px p95 absolute; conservative centerline uncertainty ±9.59 px.

## Corrected attempt results

| Attempt | Old chest/bar | Corrected chest/bar | Change | Corrected sensitivity envelope |
|---:|---:|---:|---:|---:|
| 1 | -0.169 m | -0.115 m | +0.055 m | -0.140 to -0.092 m |
| 2 | -0.186 m | -0.202 m | -0.017 m | -0.234 to -0.173 m |
| 3 | -0.255 m | -0.361 m | -0.106 m | -0.403 to -0.321 m |

Attempt 1 remains best. Its corrected chest-center peak is about 11 cm below the tracked bar, not 17 cm below the old fixed line; its sensitivity envelope is about 9–14 cm below. Attempt 2 is about 20 cm below and attempt 3 about 36 cm below.

The corrected bar and camera tracks are exported in `scene-track.csv` and `scene-samples.csv`; visual checks are `bar-track-contact-sheet.jpg` and `camera-bar-diagnostics.png`.