Measure planar distances with a calibrated camera

You can remove the calibration target before production measurements, provided the camera remains fixed relative to the measurement plane. The missing ingredient is usually the plane’s position and orientation, not a checkerboard that stays beside every part. This procedure is for inspection stations measuring features on a known, flat surface. You will establish that surface once, save the mapping, convert detected image points into physical coordinates, and verify when the mapping remains valid.

Separate lens calibration from the measurement plane

Camera calibration answers two different questions:

• Intrinsic calibration: How do camera coordinates become image coordinates, including focal parameters, principal point, and lens distortion?
• Plane registration: Where is the measurement surface relative to the camera, and what physical coordinates belong to its image points?

Intrinsic calibration alone does not establish the location or scale of an arbitrary production surface. Taking calibration images at roughly the production distance does not replace plane registration.

For a fixed camera and fixed surface, you can establish the plane once with a removable reference. Save its rotation and translation, or an equivalent metric homography, alongside the intrinsic calibration. Production measurements then use those saved parameters without the reference target appearing in every image.

The procedure for recovering a ground-plane pose from known points explains the geometric relationship. For a measurement workflow based on reference markers, use the related guide to camera calibration for length measurement.

Plan the geometry and image sampling

1. Define the physical plane containing the endpoints you intend to measure. Record the required distance tolerance, field of view, camera position, and expected part-height variation.

A conveyor surface, fixture face, and raised component surface are different planes. If you measure endpoints on a component’s top face, registering the supporting table is not automatically sufficient.

Choose the field of view around the inspection region, including placement variation and a margin for validation features. For a hypothetical image spanning 200 mm across 4000 pixels, the average sampling is:

This average is useful for sizing the camera, but perspective can make local sampling vary across a tilted surface. The calibrated mapping handles that variation; a single scale factor does not.

Use the camera-resolution calculator to check whether your smallest feature receives enough pixels. Sampling is only one contributor to measurement uncertainty, not a claim of achievable accuracy.

For more detail on this distinction, read calculating physical pixel size from magnification. Before ordering, compare the required sampling with the options in the camera finder.

Calibrate the camera and register the actual surface

2. Calibrate the intrinsics using multiple target views, then capture a separate reference image at the final measurement plane.

Intrinsic calibration benefits from target positions and tilts that cover the image rather than repeatedly showing the same centered view. Keep the lens focus, aperture, image dimensions, and any optical accessories representative of production. Follow the calibration software’s documented capture requirements.

Inspect the detected corners before accepting the calibration. An apparently successful calculation can still contain incorrect correspondences or poor edge coverage. If detection is unreliable, work through checkerboard corner-detection checks before collecting more images.

After mounting the camera in its final position, place a reference with known dimensions on the measurement plane. Define its origin and axis directions explicitly. For example, the target’s horizontal axis might define increasing fixture coordinates rather than increasing image columns.

A planar homography requires at least four non-collinear point correspondences. Prefer additional well-distributed points so you can assess residuals and reduce sensitivity to individual localization errors. A small reference in one corner gives weak evidence about accuracy at the opposite edge.

Use the reference to estimate the plane pose, or fit a homography between distortion-corrected image points and known physical coordinates. The reference dimensions determine the output units: dimensions entered in millimetres produce coordinates in millimetres.

Save the intrinsic parameters, distortion model, plane mapping, image dimensions, image-coordinate convention, and target units together. Also save a reference image and the lens settings. These records make later fault diagnosis much easier.

Convert detected pixels into physical coordinates

3. Detect each endpoint, correct its image coordinates for lens distortion, and map it onto the registered plane before calculating distance.

For a plane defined as Z = 0, the ideal pinhole relationship can be written as:

Here, K is the intrinsic matrix, R and t transform plane coordinates into camera coordinates, and r1 and r2 are the first two columns of R. The image coordinates u and v must be undistorted and expressed using the intrinsic matrix used to construct H.

Some software instead returns an image-to-plane mapping directly. Check the documented direction before applying an inverse. Do not invert a matrix merely because another example does so.

If undistortion changes the image origin, crops the image, or produces a new intrinsic matrix, use that same coordinate system throughout the mapping. Mixing raw-image points with a rectified-image transform produces systematic errors. The guide to real-time image undistortion covers the surrounding pipeline considerations.

You do not need to resample the entire image just to measure a few endpoints. Correcting point coordinates can avoid an extra image interpolation step, provided your software supports the required distortion model.

After transforming both endpoints, calculate their Euclidean separation in plane coordinates:

Do not apply this formula directly to image-pixel coordinates and then multiply by an average scale when the plane is tilted.

Remove the target and validate independently

4. Remove the registration target without disturbing the station, then measure an independent reference across the working region.

Use known distances that were not all used to fit the mapping. Check short and long spans, both principal directions, and positions near the image boundaries. For an example inspection region, validation spans might include 20 mm, 100 mm, and 180 mm. Choose actual lengths that represent your application.

Separate repeatability from dimensional bias. Capture a stationary reference repeatedly to evaluate image noise and endpoint detection. Then reposition it through the inspection region to reveal scale, distortion, and plane-registration errors.

At an average sampling of 0.05 mm/pixel, an endpoint-localization allowance of 0.4 pixel corresponds to 0.02 mm per endpoint. Adding both endpoint allowances conservatively gives 0.04 mm for their contribution to distance error. This estimate excludes reference uncertainty, plane-height error, residual distortion, and mechanical movement.

Set acceptance limits from the process tolerance rather than borrowing a universal calibration threshold. A low reprojection residual is useful, but it does not prove dimensional accuracy throughout the working area. Read camera extrinsic accuracy beyond reprojection error when planning those checks.

Also test the normal endpoint detector, not just manually selected target corners. Reflection, defocus, and threshold changes can shift the detected boundary even when the geometric mapping is correct.

Choose camera and lens candidates for stable measurements

Select area scan cameras by sampling, shutter behavior, acquisition timing, and integration requirements. A stationary inspection can have different needs from a continuously moving conveyor.

The LS-CMR-0053 is a GigE camera with a C-mount and rolling shutter. It is a candidate to evaluate where the part remains stationary during acquisition.

Rolling-shutter capture can distort moving geometry because different rows are exposed at different times. If the part moves during acquisition, evaluate shutter behavior and timing before trusting a dimensional result.

Select machine vision lenses for sensor coverage, mount compatibility, working distance, field of view, and edge sharpness. The lens finder helps narrow candidates, but calibration cannot restore detail that the optics never resolved.

The LS-FFL-0567 is a C-mount lens candidate. Verify its coverage, focus range, and field of view against the selected camera and station geometry.

These parts are starting points for evaluation, not a validated measurement package. Confirm mechanical compatibility and test the intended field before committing to a station design. Once calibrated, secure the focus and aperture adjustments against unintended changes.

Recognize height errors and choose stereo when needed

5. Confirm that production features remain on the registered plane, and invalidate the mapping whenever the geometry changes.

A single camera maps each image point onto the assumed surface. It does not independently discover whether the observed feature sits above that surface.

For a camera looking normally at the plane, a simple pinhole estimate shows the scale error caused by a raised feature:

This estimate assumes both endpoints share the same height and the camera faces the plane normally. With a tilted view or unequal endpoint heights, the error also depends on position and geometry.

Stereo is a different measurement approach, not simply a way to avoid placing a target during production. It requires calibration of both cameras and their relative geometry. Measurements are usually expressed in a camera-attached coordinate frame unless you register another frame. Start with stereo calibration from paired target images if depth is genuinely required.

Re-register the plane after relative camera-to-surface movement. Recheck intrinsic calibration after focus changes, lens replacement, or other optical changes. Cropping, resizing, binning, and altered image origins require compatible parameter updates or a validated recalibration workflow.

Commission the station with a short acceptance checklist

Before relying on production measurements, record the following:

• Plane definition: Identify the physical height containing the measured endpoints.
• Coordinate convention: Record the origin, axis directions, units, and mapping direction.
• Acquisition settings: Preserve image dimensions, crop offsets, focus, and relevant processing settings.
• Independent validation: Check several known distances across the entire inspection region.
• Repeatability: Measure a stationary reference repeatedly using the production detector.
• Height sensitivity: Test the expected range of feature heights rather than assuming it is negligible.
• Change control: Define which mechanical, optical, and software changes trigger verification or recalibration.

Keep a removable check reference available for commissioning and scheduled verification. It does not need to remain in the production scene.

Start with the camera configuration workflow once you have the field of view, tolerance, and height limits. If the plane or error budget remains uncertain, send those requirements through applications support before selecting hardware.