Industrial depth camera validation for moving inspection

A depth camera can produce convincing stationary images and still fail when the inspection head moves. Exposure, depth reconstruction, bracket movement and delayed processing can each introduce different errors. For a robot-mounted inspection station, validate the complete acquisition chain rather than selecting from image quality alone. This checklist helps you define acceptance limits, test representative motion and record the evidence needed before committing to a sensor or mounting arrangement.

Define acceptance limits before choosing hardware

Choose a depth camera only after representative moving trials meet your inspection requirements. A usable image is not necessarily a usable measurement, and a successful connection says little about data age.

Start with the feature you must measure, its allowed error and the motion present during acquisition. Separate stationary measurement performance from movement-related degradation.

• [ ] Define pass and fail separately for every critical feature.
• [ ] Include minimum, typical and maximum operating conditions.
• [ ] Decide what happens when depth is missing or a deadline expires.

Use the 3D scanning method selection guide before comparing candidates with fundamentally different acquisition methods. Browse 3D sensors through the camera finder only after those requirements are clear.

Check motion during the complete acquisition interval

Estimate image motion before running the moving trial. For approximately planar surfaces and lateral translation, object-space sampling gives a useful first screening calculation.

Suppose the required field width is 256 mm across 1,280 horizontal samples. The nominal spacing is:

256 mm / 1,280 = 0.20 mm per sample.

At a transverse speed of 500 mm/s, a trial blur limit of 0.5 sample gives:

0.5 × 0.20 mm / 500 mm/s = 0.00020 s = 0.20 ms.

These are hypothetical application values, not specifications for any listed product. The camera resolution calculator below is a sampling aid; its calculation does not establish depth accuracy.

• [ ] Test motion parallel to the image plane and toward the camera.
• [ ] Include wrist rotation, acceleration and direction changes.
• [ ] Compare stationary and moving measurements using identical settings.

The estimate does not account for rotation, rolling readout or reconstruction across multiple exposures. A hypothetical acquisition spanning 8 ms permits 500 mm/s × 0.008 s = 4 mm of travel, even if each individual exposure is short.

Confirm the complete depth acquisition interval from manufacturer documentation. Short exposure alone does not prove suitability for continuous motion. If movement creates missing regions, use the moving-object depth-shadow diagnostic to distinguish acquisition effects from later processing.

Measure valid depth where the decision happens

Whole-frame validity can hide failure on the feature you actually inspect. Evaluate each critical region before filtering, interpolation or point-cloud smoothing.

For example, 9,800 valid samples among 10,000 expected samples gives 98% validity. That percentage still fails if the missing samples form a continuous gap across the measured edge. Record both the fraction and the location of missing depth.

• [ ] Test representative dark, reflective and low-texture surfaces where they occur in your parts.
• [ ] Include edges, recesses and tilted surfaces in the evaluation.
• [ ] Record measurement bias and repeatability separately.
• [ ] Save raw depth, validity information and the final inspection decision together.

Use a reference whose measurement uncertainty is small enough for the decision you need to make. Repeatability on a fixed part cannot reveal a consistent distance bias.

Compare acquisition settings without changing several controls simultaneously. Increasing exposure may recover samples while making moving measurements worse. Aggressive filtering may produce a smoother surface while removing a real defect. Keep raw-data checks separate from the processed inspection score.

Verify usable coverage throughout the motion envelope

Do not infer usable depth coverage from the displayed intensity or color image. Stereo overlap, distance limits and processing crops can reduce the region that provides valid measurements.

The depth-image coverage explanation is the closest companion to this check. Establish the physical measurement envelope with the working-distance and positioning guide, then verify it using actual depth.

• [ ] Place the reference at the nearest and farthest required positions.
• [ ] Repeat at each image corner and along the inspection path.
• [ ] Include part-position variation and mounting tolerances.
• [ ] Check that the gripper or tool does not obscure the required surface.

Allocate coverage margin explicitly. If a feature can shift by 6 mm and mounting uncertainty contributes another 2 mm in the same direction, their conservative sum is 8 mm. A 10 mm edge margin would leave only 2 mm for other variation. This is a geometric allowance, not evidence of measurement quality near that edge.

Freeze output size, crop and depth-processing settings during qualification. If these settings change, repeat the coverage checks rather than assuming the earlier result still applies.

Separate mounting instability from sensor error

Begin with a stationary reference and a controlled mounting arrangement. Then repeat the measurement after representative robot moves. This separates sensor repeatability from bracket movement and incomplete settling.

The manual tilt stage is a bench-test candidate for controlled angle changes. Confirm its mounting compatibility and suitability for the test before ordering; it is not presented here as a robot-mounted component.

• [ ] Compare a stationary baseline with measurements immediately after motion.
• [ ] Repeat at progressively longer settling intervals.
• [ ] Check different wrist orientations for gravity-dependent bracket movement.
• [ ] Observe cable forces throughout the permitted movement envelope.
• [ ] Recheck a fixed reference after repeated motion cycles.

Select fixtures from camera stands using the accessory finder or stand configuration step. Bench fixtures and production mounts serve different purposes; verify each against its actual loads.

For a robot-mounted installation, follow the robot manufacturer's documented requirements for payload, center of gravity, inertia and cable routing. Perform movement tests within the approved safeguarding procedure. Use the extrinsic calibration accuracy guide to plan verification beyond a quick image check after a mechanical change.

Budget transfer load and data age separately

A supported interface does not mean zero latency. Measure exposure timing, transfer completion, processing completion and decision delivery as separate events.

For an uncompressed depth stream, the payload estimate is:

A hypothetical 1,280 × 720 stream with 2 bytes per sample at 30 frames/s requires:

1,280 × 720 × 2 × 30 = 55,296,000 bytes/s, or about 55.3 MB/s.

That is approximately 442 Mb/s before protocol overhead, metadata or additional image streams. Use the actual payload representation; a transferred point cloud can require substantially more data than a depth map.

Data age has a different consequence. At 500 mm/s, an uncompensated timing error of 20 ms corresponds to:

500 mm/s × 0.020 s = 10 mm of position error.

An allocation of 1 mm would allow only 1 / 500 = 0.002 s, or 2 ms, of uncompensated timing error. Longer processing latency may still be workable when measurements are correctly associated with capture-time pose and the decision deadline is met.

• [ ] Measure latency distributions, not just averages.
• [ ] Track frame identifiers, dropped frames and queue growth.
• [ ] Validate timestamp meaning and clock relationships.
• [ ] Test with production logging and other normal workloads enabled.
• [ ] Reject expired data rather than treating it as current.

The timestamp-to-result matching guide covers association checks. The camera-to-robot compatibility checklist addresses the surrounding interfaces.

Shortlist parts without assuming application performance

Catalog type identifies a candidate technology, not a successful moving-inspection solution. Request the missing acquisition, timing and measurement information before committing to a trial.

The depth-sensing camera is listed as a ToF sensing camera. Treat it as a candidate for evaluation, not as evidence of motion tolerance, accuracy or robot compatibility.

The projected structured-light camera provides another acquisition-method candidate. Confirm whether its required capture sequence is compatible with your movement and cycle requirements.

For processing, compare vision controllers against the required camera interface, software compatibility and measured workload. The vision controller candidate belongs on that shortlist only after those checks; no processing-rate claim is established here.

Before ordering, request confirmation of:

• [ ] Supported output formats, trigger behavior and timestamp definitions.
• [ ] Exposure and complete depth acquisition intervals.
• [ ] Measurement performance over the required distance range.
• [ ] SDK, operating-system and licensing compatibility.
• [ ] Mechanical drawings and documented mounting requirements.

Use the system configurator to organize the parts, but retain test evidence as the basis for acceptance.

Run a repeatable sign-off and know when to change the task

Run the same reference parts through stationary, nominal-motion and worst-required-motion conditions. Keep a record of camera settings, software versions, mounting arrangement and timing configuration.

A pilot run of 1,000 cycles with no observed failure is useful evidence, but it does not establish a guaranteed failure probability. Choose trial duration and sample diversity according to the consequences of a missed inspection.

• [ ] Every critical feature meets its measurement and validity limits.
• [ ] Coverage passes throughout the required position envelope.
• [ ] Mounting movement remains within its allocated error.
• [ ] Maximum observed data age meets the decision deadline.
• [ ] Missing frames, invalid depth and communication loss produce a defined response.
• [ ] Restarting the system restores the qualified configuration.

Continuous moving capture is not always the right answer. If acquisition motion dominates the error, test stop-and-capture operation with a measured settling interval. If coverage is inconsistent, change the viewpoint or inspection path. If depth adds no useful discrimination, a simpler imaging method may be preferable.

Take the acceptance table, representative parts and recorded motion conditions to applications support before selecting the final hardware. A documented failed test is more useful than an attractive image without a measurable acceptance criterion.