Camera calibration

Fit a geometric imaging model#

Intrinsic parameters describe how an ideal camera maps rays to pixels. Common values include focal length and principal point. A distortion model accounts for systematic bending or displacement caused by the lens. Extrinsic parameters describe the camera pose relative to the calibration target for each image.

Zhang's calibration method observes a planar pattern at several orientations. It produces a closed-form initial solution, models radial distortion, and then applies nonlinear maximum-likelihood refinement. The target or camera may move without the motion being known beforehand.

Calibration supports metric robot vision#

Calibrated intrinsics are inputs to the perspective-n-point problem, stereo reconstruction, visual odometry, and measurements derived from image geometry. A robot may also calibrate several cameras so their observations share a consistent coordinate system.

Camera calibration differs from hand-eye calibration. The first estimates a camera's imaging parameters and its pose relative to a target. The second estimates the fixed transformation between a camera and a robot frame from coordinated camera and robot motions. A deployment may require both.

A calibration belongs to a specific setup#

Changing focus, zoom, aperture, resolution, image crop, or lens mounting can invalidate previously estimated parameters. Temperature and mechanical shock can also move a camera or alter its optics. The chosen model may leave residual error when a lens has distortion that the model cannot express.

Low reprojection error is useful but not sufficient on its own. A small image-space residual concentrated near the calibration pattern may hide larger geometric error elsewhere in the field of view or at task-relevant depths. Robots should validate calibration with independent images and the same capture settings used in operation.