- Python 60.5%
- HTML 39.5%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
1779a78810
Plain ArUco for monocular object localisation, so the ChArUco board is out of the picture and the whole dictionary is free: 12 markers at 40 mm, ids 0-11, three per row on A4. Verified by round trip at 39.93 mm against 40.00 nominal. calib_cam.py calibrates the arm camera, which is the hard blocker -- estimatePoseSingleMarkers needs K and dist, and the gimbal's deg/us is unpinned for the same missing reason. The board is propped and stationary while the GIMBAL does the waving. A propped board cannot be motion-blurred and the sweep is repeatable, which is strictly better than hand-holding. It hunts coarsely for the board first, then sweeps a fine grid around wherever it found it. Captures at 1296x972 rather than 640x480: corner localisation error is roughly constant in pixels, so the wider frame is directly a more accurate calibration, and marker decoding benefits later too. Detection is findChessboardCornersSB with NO legacy fallback, deliberately -- calibrate.py's lesson that the two detectors disagree on corner order applies verbatim here. Documented in the docstring: re-propping the board between sweeps is not optional. Rotating the camera about its own centre with the board fixed gives views related by a homography, which constrains intrinsics only weakly. The sweep buys image-position coverage (pins distortion); re-propping buys pose diversity (pins focal length and principal point). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
| calib | ||
| gimbal | ||
| .gitignore | ||
| aruco-explainer.html | ||
| CLAUDE.md | ||
| pi5-rail-telemetry.html | ||
| README.md | ||
stereo-vision
Stereo depth perception for an indoor wheeled rover: two IMX296 global-shutter cameras on a 68 mm baseline, driven by a Raspberry Pi 5 with a Hailo-8 accelerator.
Status: hardware verified end to end; calibration tooling written and validated; no calibration has been produced yet. Producing one is the next step and blocks everything downstream — SGBM and every learned matcher alike need rectified pairs.
Layout
| Path | |
|---|---|
CLAUDE.md |
hardware survey, rig geometry, prior art, and the traps. Read first. |
calib/ |
calibration target generator, headless capture, and calibration + checks |
pi5-rail-telemetry.html |
how to read the Pi 5 PMIC's rails and the rest of the board telemetry |
Calibrating
python3 calib/make_target.py --page letter # or --page a4
# print at 100%, glue FLAT to something rigid, then MEASURE it
ssh pi@freddie
cd ~/stereo-calib
python3 capture_pairs.py --out calib_data --pairs 24
python3 calibrate.py --in calib_data --span-x <measured> --span-y <measured>
Pass the board spans you physically measured, never the nominal values — print scaling goes straight into your depth scale. The two axes are taken separately because printers scale the carriage and paper-feed directions differently.
calibrate.py self-checks four things: reprojection error, rectified vertical
residual, baseline against the known 68 mm, and left/right eye ordering via the
sign of T[0]. make_synthetic.py generates a known-truth dataset so the
pipeline can be validated without touching hardware.