Camera Calibration and Why It Drifts

A camera does not measure millimetres. It measures pixels, and the conversion between them exists only because somebody established it once by imaging a target of known dimensions and solving for the mapping. That mapping is not a property of the software. It is a physical relationship between a lens, a sensor, a mounting bracket and the plane the part sits in, and every one of those is a mechanical object that shifts a little over months and can shift a lot in a single maintenance intervention.

What a calibration actually establishes

• Scale. How many millimetres correspond to one pixel at the part plane. This is the number that turns a measured 340 pixels into a dimension.
• Lens distortion. Real lenses bend straight lines, most visibly at the edges of the field. A calibration models this and unwinds it, and without it a dimension is accurate at the centre and progressively wrong towards the corner.
• Perspective and the optical axis. Unless the camera looks exactly perpendicular to the part plane, there is a projective relationship between the image and the plane. Calibration recovers it, which is what allows a measurement anywhere in the frame to be meaningful.
• The working distance reference. The calibration is valid at one distance. A part sitting closer or further away is imaged at a different magnification — this is the effect telecentric lenses exist to remove.

Why a calibration drifts

Drift is not a software event. It is the accumulation of small mechanical and thermal movements, each individually negligible, that together move the mapping by more than the tolerance allows. A workshop that is cool in the morning and warm by mid-afternoon expands the bracket, the camera body and the part fixture by different amounts, because they are different materials. Vibration gradually loosens a fastener that was torqued correctly on installation. A camera bracket that was rigid enough for a static test flexes slightly when a heavy cable is dressed differently after a repair.

The intervention that catches people out is the lens. When a lens is damaged, the natural response is to order the same model and fit it, expecting the calibration to hold. It will not. Two units of the same lens model differ in principal distance and in the exact form of their distortion, and the mounting introduces a fresh tilt and a fresh working distance. A same-model swap is a new optical system. It requires a fresh calibration and, just as importantly, a fresh acceptance run, because the images it produces are not the images the model was validated on.

Writing the re-verification cadence into acceptance

A cadence that lives only in the integrator head is not a cadence. It belongs in the acceptance document, with the reference artefact identified, the measurement procedure written down, the limit stated as a number, and the action defined for exceeding it. A practical starting set of triggers is: at every production start, after any mechanical intervention on the camera, the lens, the bracket or the fixture, after any temperature excursion the workshop would consider unusual, and otherwise on a monthly interval. For dimensional work, tighten all of those; for presence and absence checks, most of them can be relaxed.

The criterion itself needs to be chosen with the same care as the inspection tolerance, and it must be tighter. If the part is judged against a tolerance of plus or minus 0.1 mm, a calibration that is allowed to drift by 0.1 mm has consumed the entire budget before the part is measured. A common and defensible convention is to hold the calibration within one tenth of the inspection tolerance. This is the same reasoning that governs how acceptance criteria for the whole station should be worded, which is covered in acceptance criteria; the drift limit is simply the part of that document that outlives commissioning. Model drift on the algorithm side is a separate line that also has to be monitored, described in model drift and recalibration.