Skip to content

Camera intrinsics

Focal length, principal point and distortion from a recording of a calibration board, written as a Kalibr camchain.yaml:

dedrunk intrinsics camera board_recording --camera /cam0/image_raw --target target.yaml

Repeat --camera for several cameras. Each is fitted on its own, so they need not see the board at the same time.

Target

Kalibr target files are read as they are.

Works with a partly visible board, which suits wide-angle lenses.

target_type: 'aprilgrid'
tagCols: 6
tagRows: 6
tagSize: 0.088      # tag edge, m
tagSpacing: 0.3     # gap, fraction of tagSize

Counted in inner corners. The whole board must be visible.

target_type: 'checkerboard'
targetCols: 7
targetRows: 5
colSpacingMeters: 0.04
rowSpacingMeters: 0.04

Kalibr has no ChArUco target. The keys follow OpenCV.

target_type: 'charuco'
squaresX: 8
squaresY: 6
squareSize: 0.04
markerSize: 0.03
dictionary: 'DICT_4X4_50'

Measure the printed board. A scale error in the square size does not change the intrinsics, but carries into every metric result that uses the board.

Recording

One to two minutes. Bring the board into every part of the image, including the corners, tilt it by 30–45° about both axes, and vary the distance. Avoid motion blur.

Lens model

--model pinhole-radtan (default) is OpenCV's and ROS's plumb_bob with k3 = 0, for lenses up to about 90°. --model pinhole-equi is the Kannala-Brandt fisheye model. If the reprojection error stays above a pixel with good images, try the other one.

Output

board found in 183 of 241 sampled images
60 views, 2076 corners (4 dropped as outliers), model pinhole-radtan
focal length:    fx 519.37 ± 0.58   fy 517.29 ± 0.66 px
principal point: cx 323.72 ± 0.56   cy 234.86 ± 0.54 px
distortion:      k1 -0.25014 ± 0.00133   k2 +0.08076 ± 0.00246   p1 +0.00086 ± 0.00011   p2 -0.00059 ± 0.00011
reprojection:    0.152 px RMS
worst views:     0.24 px at 8.0 s, 0.23 px at 0.5 s, 0.22 px at 7.0 s
board tilt:      up to 38° against the image plane
coverage:        98% of the image

Uncertainties are one standard deviation. Corners far beyond the typical error are dropped as outliers. The worst views carry their time, so --from and --to can cut them. Coverage is counted on an 8 × 6 grid. Images are sampled every camera_intrinsics.sample_interval, and at most camera_intrinsics.max_views views, spread over the image, enter the fit.

--bundle board.dedrunk saves the views for the GUI.

Use

The extrinsic commands take the camchain with --intrinsics. It overrides CameraInfo unless a camera is given as IMAGE=CAMERA_INFO:

dedrunk extrinsics lidar-camera recording --camera /cam0/image_raw --intrinsics camchain.yaml

Kalibr's kalibr_calibrate_imu_camera reads the same file as --cams.