A depth frame that comes back wavy, torn or doubled points at one mechanism on an indirect time-of-flight (iToF) camera: the target moved while the sensor was still collecting its four phase samples. If you specify 3D cameras for robot guidance, logistics or pick-and-place, you need to know how large that error gets and which sensor architecture keeps it small. You get the capture-window arithmetic here, a worked conveyor example and the decision that follows from it: belt or arm speed and depth tolerance go in, a maximum capture window comes out. With that number you can judge any iToF camera on its own timing instead of by sample images.
What a motion artifact looks like on a depth frame
Random depth noise and motion artifacts look alike in one frame and behave differently over time. Noise stays within a few millimeters of the true surface, scatters, and averages out across frames. A motion artifact is systematic: it grows with target speed, repeats identically when the motion repeats, and concentrates where the scene changes.
Typical symptoms:
- Edges of moving parts stretch or tear in the direction of travel, because samples taken early report the near position and samples taken late report the far one.
- Flat surfaces boil: the depth reading wobbles frame to frame even at constant speed.
- A smear or halo trails a fast part, with pixels holding depths between the part and the background it covered.
- Confidence values drop along moving edges, so the confidence map itself sketches where the artifact lives.
Expect it wherever depth is captured during motion: robot guidance while the arm still moves, conveyor scenes without encoder synchronization, and cameras riding a vehicle or gantry.
Four phase samples mean four capture times
An iToF camera emits amplitude-modulated near-infrared light and measures the phase shift of the returned modulation. One depth value needs four correlation samples, taken at demodulation phases 0°, 90°, 180° and 270°. Depth follows from the phase, and the unambiguous range follows from the modulation frequency:
with c = 3 × 10^8 m/s. Industrial iToF sensors typically modulate in the tens of MHz, which sets the working range:
| Modulation frequency | Unambiguous range |
|---|---|
| 20 MHz | 7.5 m |
| 50 MHz | 3.0 m |
| 100 MHz | 1.5 m |
The phase-to-distance walkthrough is in indirect time of flight: how phase becomes depth, and range folding in choose a time-of-flight unambiguous range for depth capture.
The artifact starts here: the four samples land at four different times. The phase, and with it the depth, belongs to no single instant — it mixes four of them. The elapsed time from the first sample to the fourth, the capture window, is all the time the scene has to change the result. A conventional sensor exposes a phase frame, reads it out, then exposes the next, so the window carries whole readout periods. A sensor that stores all four phase images on chip exposes them back to back and reads out afterwards, so the window holds only the four integrations.
The capture-window rule for depth error
Motion converts to depth error through one product: radial target speed times capture window.
A target at 1 m/s smears roughly 1 mm of depth per millisecond of window. Samples taken early hold the near distance, samples taken late hold the far distance, and an edge between them stretches across the span. Only the velocity component along the camera's line of sight maps directly onto depth; lateral motion shears edges instead, and accelerating or rotating parts combine both effects. Size the budget on the radial component and treat the product as the span an edge can smear across.
Worked example: conveyor parts at 500 mm/s against a ±1 mm tolerance
Inputs: a unit-load conveyor at v = 500 mm/s (0.5 m/s), a gripper that tolerates ±1 mm of depth error, a 50 MHz modulation frequency so the 3 m unambiguous range covers the scene, and two example window values — 10 ms for a sequential readout architecture, 1 ms for on-chip four-phase storage. Read the real figures from the timing section of the sensor datasheet you are evaluating.
Rule in words: the part must not travel farther inside the capture window than the tolerance allows.
Arithmetic:
- Window budget: ΔT_max = 1 mm ÷ 500 mm/s = 0.002 s = 2 ms.
- Sequential readout, ΔT = 10 ms: smear = 500 mm/s × 10 ms = 5 mm. Five times the tolerance — fails.
- On-chip storage, ΔT = 1 ms: smear = 500 mm/s × 1 ms = 0.5 mm. Half the tolerance — passes with a factor of 2 in hand.
- Frame check: a 1 ms burst sits easily inside a 30 fps depth period of 1000 ms ÷ 30 = 33.3 ms, while the belt covers 500 mm/s × 33.3 ms = 16.7 mm between depth frames — that figure budgets robot tracking, not the artifact.
Decision: at 500 mm/s, only a capture window of 2 ms or less holds ±1 mm. On-chip four-phase storage reaches the 1 ms class here; a 10 ms sequential design fails by a factor of 5. If no camera on your shortlist gets under 2 ms, stop the belt for the shot or move to a per-line, encoder-synchronized method.
What on-chip four-phase storage changes
Charge-domain storage nodes hold all four phase images completely on the chip before any readout. Three consequences matter for moving scenes:
- The four integrations run back to back, so the capture window collapses to their sum — the millisecond class instead of the ten-millisecond class of the worked example.
- Readout and phase sorting happen after the burst, outside the window; readout time can grow with resolution or frame rate without widening the motion error.
- The camera sorts the data on chip and streams depth, intensity and confidence directly, so the host merges no phase frames. Host load drops and sustained frame rates rise.
A representative industrial iToF stream runs 1.2 MP of depth at 30 fps — a 33.3 ms period that carries the burst plus full readout with room to spare. Sensor research pushes the window lower still: pseudo four-tap demodulation samples overlapping phase pairs within a single integration, trimming the effective span below the plain four-sample sequence.
Two problems remain untouched: blur inside a single long integration, which is ordinary exposure blur, and frame-to-frame latency, set by the depth period. The first is covered in ToF camera exposure time and frame rate trade-off explained, the second in depth camera frame rate drops: exposure settings cap the FPS.
For a depth stream in a robot or logistics cell, the MV-DT01WMU is the shop's ToF sensing camera.
- Type
- Depth / ToF camera
Choosing by target speed: ToF or a different 3D method
Work the budget backwards from tolerance and speed, then pick the technology:
- Part parked, camera shooting between motions: the window is irrelevant, and any iToF camera qualifies; on-chip designs simply add headroom.
- Slow radial motion up to about 100 mm/s at ±1 mm tolerance: ΔT_max = 1 mm ÷ 100 mm/s = 10 ms, which even sequential readouts meet. Standard iToF is fine.
- Continuous belts and moving robot cells at 300–1,000 mm/s: you need a 1–3 ms window, which means on-chip four-phase capture — the case the worked example settles.
- Material that never stops: per-line measurement locked to an encoder. The MV-HLVS-16K100-L60 line-scan 2.5D system samples at 120 kHz across a 600 mm emitting area, pinning each line of depth to the instant it passed.
- Type
- 2.5D inspection
- Parked parts needing full-frame surface detail: projected structured light runs its pattern sequence at a still scene. The MV-DPS200P-02 projected structured-light 3D camera fills that role.
- Texture-poor or reflective scenes on a moving rig: the MV-DLS300P-04 laser galvanometer stereo camera combines steered laser texture with stereo views captured in the same instant, freezing motion down to the exposure time.
- Type
- Stereo / structured light
- Type
- Stereo / structured light
If the cell itself is still undecided, 2D or 3D camera for conveyor pick and place settles the system question first, and 3D bin picking for flat parts covers the hardest picking case.
Checks and limits on a moving scene
Five checks keep the window arithmetic honest on your own cell:
- Compute the window from the sensor's own timing, not from the fps figure. Two 30 fps cameras can differ by an order of magnitude in phase-to-phase spacing.
- Mask by confidence during motion: pixels whose confidence drops at speed are the artifact speaking. The 3D camera confidence image bitflags guide shows how to read and apply the flags.
- Watch the unambiguous range at speed: a scene beyond it folds depth back, and the usual cure — a second modulation frequency — adds its own four-sample set and its own window.
- Keep each single integration shorter than tolerance ÷ speed, so per-frame blur stays inside the same budget; exposure time vs shutter speed for moving-part inspection gives the blur arithmetic.
- Re-shoot after geometry or parameter changes instead of trusting cached depth; why a 3D camera needs two captures after a parameter change explains the double capture.
Positioning follows ordinary rules either way: the 3D camera working distance and positioning guide covers standoff and placement, and if the camera rides the arm, 3D camera robot integration compatibility checks come before the first cycle.
If your speeds sit in the 300–1,000 mm/s band, shortlist iToF cameras with on-chip four-phase storage and hold each one to the 2 ms budget from the worked example. Browse the 3D sensors category, filter with the camera finder, or start from the camera configurator step by step; related guides live on the 3D topic page, and sizing questions reach the team through the contact page.
Questions
How fast can a target move before an iToF depth frame tears?
Use the window budget: maximum window = depth tolerance ÷ radial speed. At a ±1 mm tolerance the budget is 2 ms at 500 mm/s, 1 ms at 1,000 mm/s and 0.5 ms at 2,000 mm/s. A sensor whose four phase samples all land inside that window keeps the smear at or under the tolerance; a slower capture needs a stopped part or a per-line, encoder-synchronized method instead.
Does a higher frame rate reduce iToF motion artifacts?
No. Frame rate shortens the time between completed depth frames; the artifact comes from motion inside the four-sample window, which frame rate does not touch. A 60 fps camera with a 10 ms capture window tears exactly as much as a 30 fps camera with the same window. Higher fps does improve how fresh each depth frame is for robot tracking.
Why does on-chip four-phase storage lower the host CPU load?
Without it, the host receives four raw phase frames and must align, difference and convert them per pixel before depth exists. A sensor that stores the four images on chip can sort them into depth, intensity and confidence streams before transmission, so your code reads finished depth directly. At 1.2 MP, that is roughly 4.8 million phase pixel values per frame that never reach the host.
When should I use a laser profiler instead of an iToF camera?
When the material never stops or moves faster than the window budget allows. A profiler measures one line at a time and ties each profile to an encoder position, so part motion becomes scan geometry instead of a phase error — conveyor and web inspection are its home ground. Choose iToF when parts stop or move slowly, when you want a full depth frame in one shot, or when multi-meter volume measurement matters more than edge sharpness.
Do structured-light 3D cameras suffer the same motion artifacts?
Systems that project a sequence of patterns do, because the recovered phase again mixes several capture times into one result — keep the part still while the sequence runs. Methods that gather all raw data in a single exposure, such as synchronized stereo pairs, freeze motion at the exposure time instead. Before judging any 3D camera on a moving scene, check whether its depth result needs one exposure or several.