Structured-Light 3D Inspection: Calibration and Phase Failure Modes
Treat the projector as an inverse camera and control saturation, shadow, motion, correspondence, and phase unwrapping before trusting the point cloud.
A structured-light system reconstructs a calibration board accurately but places a step on a glossy production part one fringe period too high. The point cloud is internally smooth because the phase was decoded consistently—at the wrong absolute order. Robust 3D inspection requires valid pattern capture, correspondence and calibrated geometry.
This is a vendor-neutral engineering method. The worked example is hypothetical and must be replaced by measurements from the real product, line, and risk assessment.
What you will learn
Identify the physical, optical, data, or process limit behind structured-light 3d inspection.
Convert the inspection need into measurable acceptance criteria.
Compare practical architectures and their trade-offs.
Commission the method using repeatable evidence.
Validate the final system under representative production variation.
Technical foundation
The projector is part of the geometry
Structured-light reconstruction identifies which projector coordinate illuminated a camera pixel. Calibration therefore includes camera intrinsics, projector intrinsics and their relative pose.
Codes solve different correspondence problems
Gray codes provide robust absolute regions at discrete resolution; phase-shift patterns provide finer subpixel correspondence but require unwrapping. Combined schemes trade image count, accuracy and robustness.
Every pattern image must describe one scene
Sequential patterns assume the object and optical system remain stable. Motion, flicker, rolling shutter, vibration and exposure changes corrupt the code sequence.
Related guides on this publication: 3D Machine Vision: Stereo vs Structured Light vs Time of Flight and Camera Calibration for Machine Vision Measurement and Motion Blur in Machine Vision: Calculate Exposure Time.
Engineering workflow
1. Specify the 3D evidence
Evaluate: height, flatness, volume, edge location, tolerance, range, coverage and surface class.
Why it matters: required uncertainty and density determine pattern and geometry choices.
Measure or calculate: translate tolerance into allowed bias, repeatability and invalid coverage. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: high accuracy can require more patterns and longer capture. Common failure: selecting a scanner from point count.
2. Design projector-camera geometry
Evaluate: baseline, angles, working distance, focus, depth range, occlusion and projector focus.
Why it matters: triangulation sensitivity and shadowing come from geometry.
Measure or calculate: simulate visibility and measure projected pixel footprint across depth. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: larger angle improves depth sensitivity but increases shadows. Common failure: optimising only the camera view.
3. Control pattern radiometry
Evaluate: camera exposure, projector brightness, black level, gamma, ambient light, saturation and surface reflectivity.
Why it matters: decoding assumes separable bright and dark or stable sinusoidal response.
Measure or calculate: inspect every captured pattern for clipping, contrast and shadow masks. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: lower brightness avoids saturation but weakens dark returns. Common failure: using projector display enhancements or automatic exposure.
4. Calibrate the complete volume
Evaluate: target geometry, pose diversity, focus, temperature, camera and projector parameters.
Why it matters: a good central fit may extrapolate poorly at depth and corners.
Measure or calculate: use distributed target poses and verify independent artefacts throughout the volume. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: more poses and depth planes cost setup time but expose weak geometry. Common failure: calibrating only one plane.
5. Decode with validity checks
Evaluate: Gray-code thresholds, wrapped phase, unwrapping, modulation, consistency and correspondence masks.
Why it matters: a wrong fringe order can create a smooth gross error.
Measure or calculate: inspect code confidence and compare redundant directions or neighbourhood consistency. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: strict masks reduce coverage but prevent fabricated surfaces. Common failure: filling low-modulation regions before validation.
6. Challenge production failure modes
Evaluate: motion, vibration, glossy patches, dark areas, discontinuities, contamination, thermal drift and restart.
Why it matters: factory parts violate ideal diffuse stationary assumptions.
Measure or calculate: seed controlled shifts and surfaces and measure signed 3D errors and invalid regions. Preserve settings, sample identity, operating state, and the calculation method so alternatives remain comparable.
Trade-off: shorter sequences reduce motion risk but may lower accuracy. Common failure: reporting only cloud-to-plane RMS on valid pixels.
Worked example
Hypothetical height inspection: A phase system projects 24 periods across a 1,920-pixel projector width, so one period is 80 projector pixels. Calibration maps one projector pixel to approximately 0.06 mm of lateral object position at the test plane.
Fringe period = 1,920 / 24 = 80 projector pixels
One-period correspondence error = 80 × 0.06 = 4.8 mm laterally at the test plane
The resulting 3D displacement depends on triangulation geometry, but the example shows why a one-order unwrap error is not a small subpixel mistake. Absolute-code consistency must be tested at steps and discontinuities.
Practical decision aid
| Failure | Diagnostic evidence | Control |
|---|---|---|
| Projector or camera clipping | flat-topped pattern intensities | exposure and brightness limits |
| Shadow | low modulation or no code | second view, geometry or invalid mask |
| Motion between patterns | phase inconsistency and doubled edges | shorter sequence, stop motion or snapshot method |
| Wrong phase order | smooth surface displaced by a fringe | absolute code and consistency check |
| Calibration extrapolation | error grows at field or depth edge | volume calibration and independent artefacts |
| Specular reflection | saturation or pattern reflected elsewhere | geometry, polarisation, coating or another method |
Use the table to choose the next controlled experiment, not as a universal product recommendation. A component or algorithm is acceptable only when the complete inspection cell meets pre-agreed technical and operational criteria.
Common mistakes and how to prevent them
Calibrating only the camera. projector geometry remains wrong. Prevent it by calibrating the projector-camera pair.
Leaving projector gamma uncontrolled. sinusoidal patterns distort. Prevent it by linearising or validating the transfer.
Accepting filled shadows. missing evidence becomes invented geometry. Prevent it by keeping validity masks.
Using sequential patterns on motion. correspondence combines different scenes. Prevent it by stopping, synchronising or changing method.
Checking only mean plane residual. local unwrap errors are diluted. Prevent it by testing steps, edges and full error maps.
Reusing calibration after remounting. relative pose changed. Prevent it by verification and controlled recalibration.
Validate under production conditions
Use independent planes, steps, spheres or gauge artefacts distributed across field and depth, plus representative diffuse, dark, glossy and discontinuous parts. Test ambient light, projector warm-up, camera/projector temperature, vibration, deliberate sub-pattern motion, contamination and restart. Report signed bias, repeatability, percentile 3D error, valid coverage, unwrap error rate, edge integrity, capture time and task decisions by surface and volume region.
Use representative acceptable parts, confirmed defects, boundary samples, and nuisance variation. Repeat complete part presentations rather than processing one stored image many times. Include start-up, warm-up, maximum speed, changeover, maintenance, environmental limits, communication faults, and long-duration operation where relevant.
Define acceptance criteria before reviewing final results. Preserve raw counts and denominators for false accepts, false rejects, invalid acquisitions, timing overruns, and manually reviewed cases. After release, trend leading indicators and conduct labelled audits so deterioration is detected before a customer escape.
Key takeaways
Calibrate the projector as part of the imaging geometry.
Pattern radiometry and timing are measurement inputs.
Phase precision does not guarantee the correct absolute order.
Preserve shadow and low-confidence regions as invalid.
Validate edges, steps and production surfaces across the full volume.
Follow this Hashnode blog for more practical industrial machine-vision engineering, and connect with Kivanc Ekici on LinkedIn. To explore relevant technology information, visit Algomedi.
Frequently asked questions
Why combine Gray code and phase shifting?
Gray code can provide absolute coarse correspondence, while phase shifting refines it to subpixel precision. The combination costs more pattern images.
What causes phase-unwrapping errors?
Low modulation, saturation, motion, discontinuities, reflections, noise or insufficient absolute-code evidence can assign the wrong fringe order.
Can structured light inspect moving parts?
Sequential patterns require the scene to remain stable unless the system uses special high-speed or single-shot methods. Validate the exact motion.
Why model a projector as an inverse camera?
It maps projected pixels into rays. Correspondence between projector and camera rays enables triangulated 3D reconstruction.
How should shadows be handled?
Mark them invalid and define coverage acceptance or add another view; interpolated geometry is not direct measurement.