Pairwise camera frame alignment from calibration observations

Two fixed cameras can share measurements without introducing a separate world frame. Choose one camera as the reference, estimate the other camera’s pose relative to it, and publish that relationship in the transform tree. The difficult parts are identifying exactly which frames were calibrated, keeping transform direction consistent, and proving that the alignment works beyond the images used to estimate it. This procedure covers shared-target observations, pairwise estimation, static publication, and independent checks.

What you need before calibration

You need calibrated intrinsics and distortion parameters for both cameras, a rigid calibration target with known geometry, and mounts that remain fixed after calibration. Both cameras must observe identifiable points in the same target coordinate system.

Area scan cameras can support this workflow when their image acquisition and calibration interfaces fit your software. Use the camera finder to review acquisition options rather than selecting by resolution alone.

The LS-CMR-0166 is a USB3 rolling-shutter camera candidate. For this workflow, keep the target stationary throughout both exposures and check software compatibility before choosing the camera.

Mount stability matters as much as the calibration software. Review camera stands and the accessories finder for positioning hardware. The LS-MNT-0040 is a camera-stand candidate; verify attachment compatibility, adjustment travel, and rigidity for your assembly.

This procedure estimates one pairwise relationship. If you need several cameras tied to a separate reference, use the broader shared-coordinate-frame calibration workflow.

Define the frames and transform direction

1. Write down the exact frame IDs, coordinate conventions, and direction of the transform before collecting images.

Call the calibrated camera frames A and B, and the target frame T. Use the notation T_A_B for the transform that converts coordinates expressed in B into coordinates expressed in A.

Here, t_A_B is the position of B’s origin expressed in A. It is not generally the negative of the reverse translation.

A camera optical frame commonly uses x right, y down, and z forward. A driver’s camera-body frame may use different axes. Confirm the documented convention rather than treating every frame containing “camera” as interchangeable.

Record whether your pose solver returns target-to-camera or camera-to-target transformations. The following steps assume target points are mapped into camera coordinates.

The distinction between optical and body coordinates also affects depth data. The guide to transforming between optical and camera frames explains that conversion separately.

Capture shared target observations

2. Capture paired images of a stationary target at varied positions and orientations within the overlapping view.

Keep the cameras fixed. Move only the target between captures, then wait for it to settle. Simultaneous exposures are not essential when the entire scene remains stationary between the paired images. A moving target requires synchronization appropriate to its speed and the permitted error.

As a starting capture plan, collect about 15–25 paired views and reserve another 5–10 for validation. These are planning ranges, not guaranteed sample counts. Include different distances, image locations, and target orientations rather than repeating one central, front-facing view.

Every detected point must have the same physical identity in both images. Symmetric patterns can create indexing ambiguity, especially when viewed from different orientations. Reject observations with uncertain correspondence, clipped features, saturation, or motion blur.

Target sampling provides a useful preliminary check:

This only estimates feature sampling. It does not establish pose accuracy, which also depends on optics, geometry, contrast, and point localization.

Use the target coverage and margin checks when planning image coverage. For the target itself, the calibration-target flatness guide explains why bending can corrupt an otherwise convincing fit.

Estimate the relative camera pose

3. Solve the target pose in each camera, then compose the transformations in the stated direction.

For paired observation i, estimate T_A_Ti and T_B_Ti from the known target geometry, camera intrinsics, and distortion model. Both estimates must refer to the same target origin, axes, and length units.

The composition first takes a point from B into the target frame, then from the target frame into A. Because the camera mounts are fixed, each observation should describe approximately the same camera-to-camera relationship.

The translation unit follows the target geometry. If target coordinates were entered in millimeters, the estimated translations are in millimeters. Convert them to meters before publication when the transform system expects meters.

For a production estimate, prefer a joint calibration or refinement that fits one shared camera-to-camera transform across all accepted pairs. With trusted intrinsics held fixed, the optimization can also estimate one target pose per capture while minimizing image reprojection errors in both cameras.

Independent per-pair estimates remain useful for identifying bad captures. Do not average rotation matrices element by element or average Euler angles without handling their geometry. Rotation averaging must preserve a valid rotation; joint refinement avoids treating orientation as three unrelated scalar values.

Validate direction and accuracy on held-out observations

4. Test the fitted transform using target views that were excluded from estimation.

For each held-out pair, compare the target pose observed directly in A with the pose predicted through B:

Compare target origins and orientations, not just one displayed point. Also project predicted target corners into A using A’s intrinsics and distortion model. Inspect residual patterns across the image and at different target distances.

A numerical example makes the direction explicit. Suppose a hypothetical transform rotates B coordinates by 90 degrees around A’s z axis and has translation (0.12, -0.03, 0.02) meters:

These values illustrate arithmetic only; they are not calibration values for your installation. Apply the inverse transform to recover the original point as a software check.

Set acceptance thresholds from the application. For small angular errors, lateral displacement is approximately distance multiplied by angular error:

This approximation shows why a small orientation error can matter at longer distances. Translation error and other uncertainty sources also contribute. The extrinsic calibration accuracy guide covers checks beyond a low image-fit residual.

Publish the fixed parent-child relationship

5. Publish the validated transform through your middleware’s documented static-transform broadcaster, after checking the existing tree.

A transform message normally stores the child frame’s pose in the parent frame. With parent A and child B, publish T_A_B: the same transform used above to map B coordinates into A.

For the hypothetical numerical example, the message fields would be:

header:
  frame_id: cam_a_optical
child_frame_id: cam_b_optical
transform:
  translation:
    x: 0.12
    y: -0.03
    z: 0.02
  rotation:
    x: 0.0
    y: 0.0
    z: 0.7071067811865476
    w: 0.7071067811865476

The rotation shown is a normalized quaternion in x, y, z, w field order. Check the publisher’s documented argument order; some interfaces accept Euler angles, and others expect quaternion components in a different sequence.

Do not publish this example into an operating system. Replace it with your measured, validated transform and actual frame IDs.

Before connecting frames, inspect the driver-generated tree. If B’s optical frame already has a parent, attaching another parent creates a conflicting topology. Normally you connect the two camera subtree roots and preserve each driver’s internal fixed relationships.

For camera-body frames A_link and B_link, convert the calibrated optical relationship as follows:

The first and last terms come from the documented camera frame relationships, not from assumptions about the axes. Publish the resulting body-to-body transform with the corresponding parent and child IDs.

Use one publisher for that relationship, avoid cycles, and save the calibration with the camera identities and frame definitions. The static-transform publication workflow provides related guidance on preserving direction during publication.

Transform measurements, not just their labels

6. Request the transformation into A when processing measurements originating in B, and verify that the processing node actually applies it.

Publishing a transform adds a relationship to the tree. It does not automatically rewrite every image, point cloud, detection, or pose message produced by either camera. Renaming a message’s frame ID without transforming its coordinates is incorrect.

For depth cameras, the extrinsic alignment does not remove depth-scale error or per-camera depth bias. Validate reconstructed geometry separately from image-target reprojection.

Check the transform lookup in both directions. A point transformed from B to A and back should agree numerically within floating-point tolerance. That confirms implementation consistency, but only independent physical observations confirm calibration accuracy.

When the shared-target method is the wrong choice

A rigid target observed by both cameras is the simplest route when the views overlap. It is not a complete solution for every installation.

If the views do not overlap, independently moving an ordinary target between them does not establish its motion. You need additional constraints, such as a rigid target assembly with surveyed geometry visible to both cameras, or an independently measured target motion.

If either camera moves after calibration, one static transform is no longer sufficient. A measured time-varying pose or a new calibration is required. Likewise, a flexible mount can make the relationship change during operation even though the saved transform remains constant.

If the target moves during unsynchronized capture, collecting more images does not necessarily remove the resulting systematic error. First stop the target, synchronize acquisition, or use an estimation method that explicitly accounts for timing and motion.

Final checks before deployment or ordering

Use this checklist before accepting the alignment:

• Confirm the two camera identities and exact calibrated frame IDs.
• Record the target geometry, units, intrinsics, distortion model, and image settings.
• Inspect held-out observations across the intended working region.
• Check translation magnitude and axis directions against the physical installation.
• Verify quaternion normalization and the publisher’s component order.
• Confirm that each child has one parent and each relationship has one publisher.
• Check that processing nodes transform coordinates rather than changing labels.
• Repeat the physical checks after any camera or mount adjustment.

Before ordering hardware, verify camera acquisition support, target visibility, mount compatibility, and adjustment access. Use the camera configurator to plan the imaging hardware, or send your frame diagram, target dimensions, and accuracy requirement through applications support.