Motion Blur in Machine Vision: How Exposure Time, Conveyor Speed and Shutter Type Set Image Sharpness

A moving part is not blurred because the camera is “slow”. It is blurred because, during the exposure, the image of the part travelled across the sensor by a distance you can calculate in advance: blur = velocity × exposure time. Everything else — shutter type, strobe lighting, line rate, trigger jitter — is a way of changing one of those two factors. This page gives the formula, the units, the acceptance thresholds and three worked examples, using published POMEAS camera and lighting specifications.

By the POMEAS Optics Applications Team · Reviewed by Tiger Xie · Updated 14 October 2026

1. What motion blur is, and what it is not

Motion blur is a smear along the direction of relative motion between the part and the sensor. It has a length, a direction, and a cause you can measure. That distinguishes it from the other ways an image can look soft, and the distinction matters because the fixes are different.

Effect	Cause	Signature in the image	What actually fixes it
Motion blur	Relative motion during the exposure	Directional smear; length grows with speed and exposure	Shorter exposure, strobe/overdrive light, stop the part, or line-scan
Defocus	Feature outside the depth of field	Isotropic blur; grows with distance from best focus; present on stationary parts too	Focus, aperture, working distance, or a lens with more depth of field
Lens / diffraction limit	Optical resolution ceiling	Isotropic softness that does not change with speed	A lens whose resolution matches the pixel pitch; a larger aperture where diffraction dominates
Noise-limited blur	Too little light, too much gain	Grain, not smear; edges scatter rather than stretch	More light or a longer exposure — not a shorter one

The practical test: photograph the part stationary at the same exposure, with the same illumination. If the image is sharp when the part is still and smeared when it moves, the cause is motion. If it is soft in both cases, the cause is optical or illumination-related, and shortening the exposure will only make it darker.

It is also worth separating two blur quantities that are easy to confuse. Object-side blur is the distance the part itself travelled during the exposure, in millimetres on the workpiece. Sensor-side or pixel blur is the same motion expressed on the sensor, in micrometres or pixels. Your measurement software only ever sees the second one, and the two are related by the magnification.

2. The one formula, and the units that catch people out

Let V be the velocity of the part in the object plane, measured perpendicular to the optical axis, and t the exposure time — the period over which the sensor integrates, not the frame period.

• Object-side blur (mm): B_obj = V · t
• Sensor-side blur (mm): B_img = m · V · t, where m is the image-to-object magnification
• Blur in pixels: B_px = V · t / p_obj, where p_obj = FOV / N is the object-side pixel size

The third form is the one to work with, and it contains a result worth pausing on: the object-side blur V · t does not depend on magnification, but the pixel blur does. The same 20 µm of travel that is 6.5 pixels at the wide end of a zoom is 37 pixels at the narrow end. A measurement you validated at low magnification can fail at high magnification on the same line at the same speed.

Quantity	Symbol	Unit	Note
Object velocity	V	mm/s	Multiply m/s by 1000. Use the velocity at the feature, not the average line speed.
Exposure time	t	s  (1 µs = 1 × 10−6 s)	The integration time. Not the frame period: at 30 fps the frame period is 33.3 ms, which may be a hundred times longer than the exposure.
Field of view	FOV	mm	The horizontal (or vertical) extent of the scene, consistently with N below.
Pixel count	N	px	Effective pixels across that FOV, after binning and ROI.
Object-side pixel size	pobj	mm/px (µm/px)	pobj = FOV / N = ppitch / m
Magnification	m	–	Image size ÷ object size. For a zoom lens this is the setting you dial in.
Blur	B	mm or px	State which. Reporting “blur is small” without units is how mis-specifications happen.

Four unit errors account for most of the wrong answers:

1. Using the frame period instead of the exposure time.
2. Mixing m/s with microseconds without converting both.
3. Using the sensor pixel pitch as p_obj; the two differ by the magnification, which for a zoom lens can be a factor of six or more.
4. Using a quoted “resolution” in line pairs per mm rather than a pixel count.

3. Worked example 1 — a conveyor line

The setup. A field of view 100 mm wide, digitised with 4096 pixels across it. The conveyor runs at 200 mm/s. The application is an edge position used in a dimensional check, so the target is no more than one pixel of blur.

Step 1 — object-side pixel size. p_obj = 100 mm / 4096 px = 0.0244 mm/px = 24.4 µm/px.

Step 2 — maximum exposure. Rearranging the pixel-blur formula: t_max = B_px · p_obj / V = 1 × 24.4 µm / 200 mm/s = 122 µs.

Step 3 — can the camera do it? Compare against the published exposure range of the POMEAS camera family you are considering:

Model	Shutter	Resolution / pixel	Published exposure range	Max frame rate	Interface
PMS-1209MGS	Global	4096 x 3072, 3.4 um	1 us - 10 s	67.9 fps	GigE
PMS-CCN120VP2ABSC	Global	12.6 MP, 1.1 in	1 us - 10 s	9.4 fps	GigE
PMS-CCU120P2ACSC	Global	12.6 MP, 1.1 in	7 us - 10 s	31.9 fps	USB3.0
PMS208CCUA1C	Rolling	5480 x 3648, 2.4 um	12 us - ~5.06 s	22 fps	USB3.0
PMSCCN2008VPH2	Rolling	5480 x 3648, 2.4 um	45 us - 12 s	5.9 fps	GigE
PMSMS02C1201	Rolling	12 MP, 1/1.7 in	9 us - ~30.6 s	35.9 fps	USB3.0
PMS-CCN-V0502MHSC	Rolling	2592 x 1944, 2.2 um	not published	23.8 fps	1GigE

Every model above reaches 122 µs, and the global-shutter models reach it with range to spare: their published minimum exposure is 1 µs to 7 µs. The rolling-shutter models also reach it, but see section 7 for the geometry penalty.

Step 4 — what if the line is accelerated? At 1 m/s the same limit becomes t_max = 24.4 µm / 1000 mm/s = 24.4 µs. That is still inside the published ranges, but the exposure has fallen by a factor of five. Holding the same signal level now needs roughly five times the irradiance at the part, because the charge collected is proportional to irradiance × time. If the light cannot deliver it, the alternatives are more gain (which costs signal-to-noise ratio), a narrower field of view, or a different measurement principle.

4. Worked example 2 — index motion against continuous motion

An indexing table stops the part for the exposure. If the part is genuinely stationary during the integration, V is zero and blur disappears — the cost moves to throughput, which becomes settle time + exposure + readout. Continuous motion never pays the settle time, but it always pays the blur.

Throughput comparison. With a part pitch of 30 mm:

Mode	Rate	Exposure that keeps 1 px of blur	Trade-off
Continuous, 200 mm/s, pobj = 24.4 µm	200 / 30 = 6.7 parts/s	122 µs	Blur must be managed; no settle time
Continuous, 600 mm/s, pobj = 24.4 µm	20 parts/s	24.4 µm / 600 = 40.7 µs	Needs 3× the irradiance at 200 mm/s
Indexing at 20 parts/s	set by mechanism	V ≈ 0, so blur is negligible	Pays 50 ms per part in settle plus readout; may not reach 20 parts/s at all

The term that is usually forgotten: trigger jitter. Blur is not the only speed-dependent error. If the trigger fires with a timing uncertainty σ_t, the part is at a different position at each exposure, and the positional uncertainty is V · σ_t — independent of the blur. At 200 mm/s, 50 µs of jitter displaces the part by 10 µm, which with a 24.4 µm object-side pixel is 0.41 px. Blur and jitter are separate entries in the same error budget, and reducing the exposure does nothing for the second one.

5. Worked example 3 — rotating parts, where blur varies across the field

On a rotating part, surface velocity rises linearly with radius: V = ω · r. With angular speed n in revolutions per minute, V at a radius r is 2πn r / 60, and at the rim of a disc of diameter D it is π · D · n / 60.

The setup. A 50 mm disc rotating at 300 rpm, viewed with p_obj = 24.4 µm/px, one pixel of blur allowed.

Rim velocity. V = π × 50 mm × 300 / 60 = 785 mm/s.

Maximum exposure. t_max = 24.4 µm / 785 mm/s = 31.1 µs.

Two consequences follow. First, 785 mm/s is nearly four times the 200 mm/s conveyor of example 1, so the exposure must be about four times shorter for the same pixel blur — the disc is a far more demanding target than its physical size suggests. Second, blur is not uniform across the image: a feature at half the radius moves at half the speed and is blurred by half as much. The rim, not the average, sets the specification.

The same arithmetic covers vibration. A machine vibrating at 100 mm/s blurs a 200 µs exposure by 100 × 200 × 10⁻⁶ = 20 µm, which is comparable with the object-side pixel of a large-field station — the relation used in the POMEAS calibration and stability article, which treats vibration as one of the four causes of drift on a running line.

6. How much blur can you afford?

There is no single answer, because the tolerable blur depends on what the software extracts — and the honest position is that the only reliable number comes from a gauge R&R on your own part at two or three speed levels. The figures below are the working rules we start from, not measured limits:

Application	Starting target	Reasoning
Presence / absence, orientation, coarse gauging	1 – 2 px	The decision is about a feature that is many pixels across; a two-pixel stretch rarely changes the answer.
OCR, barcode, character verification	≤ 1 px	Stroke width is only a few pixels; smearing closes the gaps the decoder relies on.
Edge location with sub-pixel centroid or Gaussian fit	≤ 1 px	Reduces edge steepness, so the extracted position scatters more between frames.
Dimensional measurement against tolerance	≤ 0.3 px, and at least five to ten times smaller than the tolerance expressed in pixels	Blur is one term among several; leaving it as the largest term wastes the rest of the budget.
Sub-pixel metrology on a repeatability-critical feature	≤ 0.2 px	The blur term must sit well below the repeatability target, or it dominates the run-to-run scatter.

Repeatibility against accuracy is worth separating here. A symmetric smear along the direction of motion tends to preserve the centroid of a well-lit, symmetric edge, so blur often does not bias the mean reading. What it does is flatten the edge, which raises the scatter, and it turns any asymmetry — a chamfer, a burr, one-sided illumination, a shadow — into a real displacement. In practice, motion blur shows up first as a repeatability and threshold-sensitivity problem, and only later as an accuracy problem.

7. Global shutter and rolling shutter: what each one costs you

Shutter architecture does not change B_px = V · t — the sensor integrates for the exposure either way. What it changes is when each pixel integrates, and that changes the geometry of the error.

Architecture	How it integrates	Consequence on a moving part	POMEAS models in the table above
Global shutter	All pixels integrate in the same time window	No skew and no banding. Short, strobe-synchronised exposures are straightforward to implement. This is the default choice for parts in motion.	PMS-1209MGS, PMS-CCN120VP2ABSC, PMS-CCU120P2ACSC
Rolling shutter	Rows are exposed and read out sequentially	A moving part is displaced progressively across the frame (skew). A strobe shorter than the row period produces a bright band rather than a uniform exposure.	PMS208CCUA1C, PMSCCN2008VPH2, PMSMS02C1201, PMS-CCN-V0502MHSC

The skew magnitude follows the same arithmetic as the blur. If the sensor takes T_ro to read a full frame, the last row is exposed T_ro after the first, so a part moving at V is displaced by V · T_ro between them. For a nominal 22 fps rolling-shutter sensor the frame period is about 45 ms, so at 200 mm/s the top and bottom of the frame see the part about 9 mm apart — a slant that is easy to mistake for a mechanical tilt in the fixture.

POMEAS states the same selection rule on the product pages. The 12 MP USB3 camera page, which uses a rolling shutter, notes that for faster motion you should use a short exposure time together with synchronised strobe lighting to reduce distortion and motion blur, and that target speed, exposure time and strobe synchronisation should all be evaluated before selection for high-speed moving objects. The 20 MP and 21 MP pages, also rolling shutter, explicitly recommend a global-shutter camera where the parts move very fast, rotate quickly or the fixture vibrates severely.

If your part geometry is what you are measuring, skew is not a cosmetic problem. A rolling-shutter skew of several millimetres across the frame changes the apparent width of a rectangular part, and unlike blur it is not symmetric, so it biases the reading rather than merely adding scatter.

8. Strobe and overdrive: what illumination can and cannot fix

A strobe shortens the effective exposure to the duration of the light pulse, and blur then follows the pulse width rather than the camera exposure. That is the single most useful lever in high-throughput inspection — but it has three hard limits, and they are the reason two-thirds of strobe projects under-deliver.

Limit 1: ambient light must be negligible. The sensor integrates for the whole camera exposure regardless of when the strobe fires. If daylight or factory lighting contributes a measurable signal, the blur comes straight back. The fixes are a dark tunnel, a shroud, or a narrow band-pass filter matched to the LED wavelength.

Limit 2: the driver sets a floor on pulse width. The published POMEAS 4-channel LED light controller drives four independent channels with 256-level dimming, RS-232 control and a 5–24 V external trigger, with a maximum trigger frequency of 1 kHz and a trigger response time of 200 µs. Read those two numbers as constraints: 1 kHz fixes the fastest strobe rate at 1000 pulses per second, so 1000 parts per second is the ceiling for one part per flash; and a 200 µs response time bounds how short a pulse the driver can produce reliably. If the calculation in section 2 asks for a 30 µs pulse, the controller specification — not the camera — is the thing to check.

Limit 3: dimming must not beat with the exposure. This is stated in the POMEAS technical reference on optical test methods: if PWM dimming is used, the PWM frequency must be at least 20 kHz or the exposure must be synchronised to the PWM, otherwise the intensity ripple appears as noise in the image. A strobe system whose driver PWM sits in the low kilohertz range can add ripple that looks exactly like sensor noise and is misdiagnosed as a camera problem.

Two more constraints come from the same reference and apply directly here: exposure and gain should be fixed for a whole sequence — a value measured at one exposure is not comparable with a value measured at another — and the brightest region of interest should sit at 70–80% of saturation, with no clipping, so that the full dynamic range is in play without the peak taking the edge centroid with it.

What a strobe fundamentally cannot do is recover information. If the exposure is long enough that the part moved far, a strobe that fires for the last 20 µs of it still integrates the part with 20 µs of travel plus whatever the ambient light contributed over the whole window. Shortening the camera exposure and shortening the pulse have to be done together.

9. A magnification sensitivity table: the same speed, four settings

This is where the pixel form of the formula earns its keep. Take the published 2/3-inch field of view of the POMEAS 6.5X zoom family and digitise it with 4096 pixels across the horizontal field. The object-side travel is the same 20 µm in every row — a 200 mm/s conveyor with a 100 µs exposure — but what the software sees is not.

Magnification	Horizontal FOV	Object-side pixel pobj	Blur from 20 µm of travel	Exposure that keeps 1 px
0.7X	12.57 mm	3.07 µm/px	6.5 px	15.3 µs
1.0X	8.80 mm	2.15 µm/px	9.3 px	10.7 µs
2.0X	4.40 mm	1.07 µm/px	18.6 px	5.4 µs
4.0X	2.20 mm	0.54 µm/px	37.2 px	2.7 µs

The numbers in the last two columns are the point. At 0.7X the same physical motion costs 6.5 pixels; at 4.0X it costs 37. And the exposure that keeps blur at one pixel falls from 15.3 µs to 2.7 µs. At high magnification the camera's minimum exposure becomes the binding constraint before the light does: of the models in section 3, only the three with a published 1 µs or 7 µs floor can reach 2.7 µs at all, while the 12 µs, 45 µs and 9 µs rolling-shutter models cannot.

(The premise of this table is a sensitivity calculation — the published 2/3-inch field of view, digitised with 4096 pixels across it. Match the lens image circle to the sensor format before you buy: a 2/3-inch format zoom lens does not cover a 1.1-inch sensor, and the depth of field falls steeply as magnification rises.)

10. When imaging is the wrong instrument

• Continuous web or strip at high line speed. A line-scan camera exposes one line at a time and is synchronised to an encoder, so the rate that matters is the line rate, not the frame rate. The POMEAS 16K line-scan range (PMS-LLC, M72 mount) exists for exactly this geometry.
• Single-dimensional measurement of a fast surface. A confocal displacement sensor sidesteps the problem entirely: the POMEAS 32 kHz single-channel controller takes AB/ABZ encoder input and pulse or level triggering, so the sampling instant is locked to position rather than to a frame. If what you need is a height or a thickness rather than an image, this is usually cheaper and faster than making an area camera keep up.
• A go/no-go decision on a known feature. A comparator on a laser or a photoelectric sensor resolves in microseconds and does not care about magnification at all.

The decision rule is simple. If the answer you need is a number of millimetres, ask whether an imaging system is the right instrument before you spend money on reducing its exposure. If the answer is a decision about a shape or a surface, imaging is the right instrument and blur is a design parameter to be budgeted, not an accident to be fixed in the field.

11. A selection procedure you can run in one sitting

1. Measure V at the feature, worst case. Maximum line speed, rim radius on a rotating part, and any acceleration peaks. Averaging over the cycle is how specifications get missed.
2. Fix the field of view and the pixel count from the smallest feature you must resolve, then compute p_obj = FOV / N.
3. Choose the allowable blur in pixels from section 6, based on what the software actually extracts.
4. Compute t_max = B_px · p_obj / V.
5. Check it against the camera's published exposure range, not just its maximum frame rate. The minimum exposure is the number that fails first at high magnification.
6. Check the illumination budget: can the light deliver irradiance proportional to 1/t at the part, without saturating the brightest region above 80% of full scale?
7. Check the shutter architecture against the geometry you are measuring, and plan the strobe and its trigger source.
8. Budget the trigger: V · σ_t for jitter, plus the trigger latency, plus the encoder quantisation if you are tracking.
9. Validate at three speed levels. Run the gauge R&R at the nominal speed, the maximum speed and one intermediate point. A measurement that passes at one speed and fails at another is a specification error, not a machine fault.

12. What has to stay fixed while you run

Every quantity in the formula is a setting, and a setting that drifts silently turns a valid measurement into an invalid one. The discipline is the same one the POMEAS optical test reference applies to lens measurement: fix the exposure and the gain for a whole sequence, log the trigger configuration with the result, and re-verify against a reference artefact after any change to speed, lighting, lens, camera or fixture. A value measured at one exposure is not comparable with a value measured at another, and on a running line the exposure is the setting most likely to be quietly adjusted by an auto-exposure routine.

If your problem is drift rather than blur, the companion page is the one on calibration and measurement stability, which breaks a running line down into its four disturbance causes — contamination, temperature, vibration and illumination ageing — and treats vibration with the same V · t relation used here.

14. Frequently asked questions

How do I calculate motion blur in machine vision?

Divide the object-side travel during the exposure by the object-side pixel size: blur in pixels = (V x t) / p_obj, where V is the object velocity in mm/s, t the exposure time in seconds and p_obj the field of view divided by the pixel count across it. Object-side blur in millimetres is simply V x t and does not depend on magnification; the pixel number does.

What is an acceptable amount of motion blur in a measurement?

It depends on what the software extracts. For presence or absence and orientation, one to two pixels is usually tolerable. For edge location with a sub-pixel centroid or Gaussian fit, keep it at or below one pixel. For a dimensional measurement against a tight tolerance, target 0.3 pixels or less, at least five to ten times smaller than the tolerance expressed in pixels. These are starting rules, not measured limits, so confirm them with a gauge R&R on your own part.

Does a global shutter camera eliminate motion blur?

No. A global shutter removes skew and banding because all pixels integrate in the same time window, but the sensor still integrates for the whole exposure, so the blur remains V x t. What a global shutter buys you is the ability to use a short, strobe-synchronised exposure without the geometry penalties of a rolling shutter.

Can strobe lighting remove motion blur completely?

It can shorten the effective exposure to the light pulse width, provided ambient light contributes almost nothing. If ambient light is significant the sensor keeps integrating and the blur returns, so a dark tunnel or a band-pass filter may be required. The driver also sets a floor: the published POMEAS 4-channel LED controller has a 1 kHz maximum trigger frequency and a 200 microsecond trigger response time, so a pulse far shorter than that needs a driver specification you have confirmed with the supplier.

How does conveyor speed affect image sharpness?

Blur is directly proportional to speed. Doubling the line speed doubles the blur at the same exposure, or forces you to halve the exposure to hold the same blur. An exposure that was validated at 200 mm/s has to be re-derived when the line is accelerated for a higher-throughput product.

Why is motion blur worse on a rotating part?

Because surface velocity rises linearly with radius: V = omega x r, which at the rim of a disc of diameter D and rotational speed n in rpm is pi x D x n / 60. A 50 mm disc at 300 rpm has a rim velocity of about 785 mm/s, nearly four times a 200 mm/s conveyor. Blur is also not uniform across the field of view, because a feature nearer the centre moves more slowly than one at the rim.

What exposure time do I need for a conveyor moving at 1 m/s?

It depends on the object-side pixel size. With a 100 mm field of view digitised by 4096 pixels, the object-side pixel is 24.4 micrometres, and one pixel of blur at 1000 mm/s allows 24.4 microseconds. At 200 mm/s the same limit is 122 microseconds. Note that holding the signal level at 24.4 microseconds needs roughly five times the irradiance of the 122 microsecond case.

Does motion blur shift the measured edge position or only soften it?

A symmetric smear along the direction of motion tends to preserve the centroid, so a centred, symmetric edge is not systematically displaced. What degrades is edge contrast, which raises the scatter of the extracted position, and any asymmetry in the edge or the illumination turns into a real bias. Blur is therefore usually a repeatability and threshold-sensitivity problem before it becomes an accuracy problem.

What is the difference between motion blur and defocus?

Motion blur is directional: it smears along the direction of relative motion and its length is V x t. Defocus and diffraction are isotropic and do not change when the part stops moving. Photograph the part stationary at the same exposure and illumination: if it is sharp when still and smeared when moving, the cause is motion.

When should I use a line-scan camera instead of an area camera?

When the part moves continuously and the whole surface has to be inspected at a resolution an area camera cannot hold at that speed. A line-scan camera exposes one line at a time and is synchronised to an encoder, so the limiting rate is the line rate rather than the frame rate. The POMEAS PMS-LLC 16K line-scan lenses with an M72 mount are built for that configuration.

Does a longer exposure always increase blur?

For a moving part, yes, because blur is directly proportional to exposure time. Long exposures remain correct for stationary parts and for low-light inspection, where the alternative of raising gain costs signal-to-noise ratio instead.

How does trigger jitter affect the measured position?

It adds a positional uncertainty of V x sigma_t that is independent of the blur. At 200 mm/s and 50 microseconds of jitter the part moves 10 micrometres between the intended and the actual exposure, which with a 24.4 micrometre object-side pixel is 0.41 pixels. Blur and jitter are separate terms in the same error budget, and shortening the exposure does nothing about the jitter.

13. Get a blur and exposure calculation for your line

Send us the field of view, the smallest feature you have to resolve, the maximum speed at the part and your tolerance, and we will return the object-side pixel size, the maximum exposure, the models whose published exposure range can actually reach it, and whether the illumination you have will support it. If the arithmetic says an area camera cannot do the job at your speed, we will say so and propose the line-scan or confocal alternative instead.

Email [email protected] · WhatsApp +86 185 9810 2007 · or use the contact form

Related reading: Calibration and measurement stability on a running line · Object-side resolution versus measurement accuracy · Magnification is not resolution · Industrial lens selection: calculations and errors · TR-001: 16 optical performance test methods · Industrial cameras · 12.6 MP global-shutter USB3 camera · 4-channel LED light controller with strobe trigger · 16K line-scan lens, M72 · 32 kHz confocal displacement controller