Shutter type is a single line on a machine vision camera datasheet that quietly decides whether a machine vision system can measure a moving part at all, and global vs rolling shutter is not a quality ranking [S5].
On a continuously moving conveyor, web, or robot-mounted rig, a global shutter freezes every pixel at the same instant, while a rolling shutter reads row by row, top to bottom, so the bottom of the frame is captured a fraction of a second after the top [S1][S2]. That row delay is the entire failure mode, and the geometric error it produces scales directly with line speed [S5].
What each shutter actually does to a frame
A global shutter starts and stops light collection on every pixel simultaneously, so the frame is a single snapshot in time [S5]. A rolling shutter exposes and reads out the sensor sequentially, line by line, which optimises bandwidth and host load but means each row sees the part in a slightly different place if the part is moving [S2][S3].
When the part is stationary this distinction is invisible, which is why shutter type is often ignored on bench setups, and why moving it to production is where the surprise shows up. With a vision controller running edge detection on a stopped part, both shutter types return the same coordinates within sub-pixel tolerance; on a 1 m/s line the same camera on a rolling shutter can return coordinates that drift with line speed [S5].
Skew is not motion blur, and the fixes do not overlap
Motion blur and rolling shutter skew are routinely conflated, but they are separate problems with separate fixes [S5]. Motion blur is the part moving during a single exposure, and a shorter exposure, with more light, fixes it for both shutter types. Rolling shutter skew is the part moving between the top and bottom rows being read, and only rolling shutters suffer from it; a shorter exposure does not fix it because the readout still takes as long as it takes, so each row can be perfectly sharp and the frame still be geometrically wrong [S5].
Concretely, a rectangle becomes a parallelogram, a circular hole becomes an ellipse, and a straight edge becomes a slanted one, with the magnitude of the error scaling with line speed [S5]. For a machine vision imaging cell that reports a dimension on a moving part, skew is a measurement error, not a cosmetic one, and the same part will measure differently at different conveyor speeds [S5].
The strobe workaround and what it actually costs
A rolling shutter can be made to behave like a global one by controlling the light instead of the sensor: open the shutter in a dark enclosure, fire a very short high-intensity flash, then finish the readout, so every row records the same instant regardless of when it was read [S5]. This is a well-known technique and is common in inspection stations where enclosing the cell and adding a strobe controller is acceptable.
Three conditions come attached, and each one is real engineering: the station has to be dark enough that ambient light contributes nothing during readout, usually requiring an enclosure; the light source has to deliver the intensity needed for a short pulse, so the vision light source is no longer a simple always-on ring; and the flash has to be synchronised to the trigger and the readout, which is one more thing to commission and one more thing that can drift [S5]. As a rough guide, OV80i-class cameras run a rolling shutter at 16 ms exposure without a strobe and about 10 microseconds with one, roughly three orders of magnitude, which is the difference between a sensor that cannot inspect a moving part and one that can [S5].
Selection criteria for on-the-fly measurement cells
The first question to answer is whether the part is moving at the moment of capture [S5]. From there, the choice between global shutter, rolling shutter with strobe, and rolling shutter alone maps cleanly to the application class.
Where the part indexes and stops in front of the camera, shutter type stops mattering because nothing moves, and the decision should be made on resolution, pixel pitch, and interface instead [S5]. Where the part moves continuously on a conveyor, a global shutter, or a rolling shutter inside a dark enclosure with a synchronised strobe, is required, and this is where the bulk of in-line vision measuring machines operate. Where the camera is mounted on a robot arm, a global shutter is required unless the arm fully settles before capture, because even a brief settle window can leave residual motion that turns into skew on a rolling sensor [S5]. Continuous web, roll, or sheet inspection is the same case as a moving conveyor: global shutter, or line scan, is the spec that survives contact with production speeds [S5].
On cost, a rolling shutter readout is less complex than a global shutter readout, which makes rolling-shutter CMOS sensors cheaper at a given resolution, and the lower transistor count also produces less heat and electronic noise, with a smaller pixel pitch that allows higher resolutions on a given optical format [S3]. The catch is that the price gap is only worth taking if the application is genuinely stationary at the moment of capture, or if the engineering budget covers a dark enclosure, a strobe controller, and a synchronised vision light source [S3][S5].
Where rolling shutter is still the right call
Rolling shutter is not a bad technology; it is the wrong tool for measuring a moving part. On static or slowly moving scenes, controlled lighting, and high-resolution inspection where the priority is pixel count per dollar, a rolling shutter is the rational choice [S3][S4]. Embedded vision modules, machine vision id readers scanning barcodes on a label after the label has stopped under the camera, and high-resolution overhead inspection of stationary PCBs all sit in this band [S2].
For 3D imaging and stereo vision, the rule flips: global shutter is the only acceptable option because the triangulation math assumes both cameras captured the same instant, and a rolling shutter on either side corrupts the depth map even when the part is moving slowly enough that the skew is invisible to the eye [S4]. The same logic applies to any multi-camera rig where frames have to be synchronised to a common trigger, which is most of the in-line cells built around a machine vision id station. A practical reference for how line-speed error budgets cascade into adjacent tooling is laid out in the cobot screwdriving accuracy breakdown; the same 0.1 mm tolerance math that bites a screwdriving cell will eat a rolling-shutter vision cell first.
Failure modes and what they look like on the line
The classic failure is a vision cell that passes acceptance test on a stopped part, then drifts out of tolerance the moment production speed ramps up. Operators see parts that pass at 0.3 m/s and fail at 0.8 m/s, with the error growing linearly with line speed, which is the signature of rolling shutter skew rather than motion blur or lighting [S5]. A second failure mode is the strobe workaround that loses its enclosure, for example a maintenance team opens the dark box to clear a jam and never re-seals it, after which the cell drifts again because ambient light is now contributing during readout [S5].
A third failure mode is using a rolling shutter for a robot-mounted inspection where the robot brief is "stop and capture" but the stop is actually a 50 ms decel, which is long enough to introduce measurable skew on a rolling sensor at typical pixel clock rates. The fix in all three cases is the same: either move to a global shutter sensor at the next camera refresh, or commit fully to the dark-enclosure + strobe architecture and treat the enclosure as a safety interlock rather than an accessory.
Sourcing and standards signals to track
Sensor-side, the global shutter CMOS roadmap is the dominant signal: Sony Pregius, Sony Pregius S, and ON Semiconductor XGS are the families most commonly specified for industrial global shutter cameras in 2024-2026 designs, and pixel pitch has dropped from 3.75 µm to 2.74 µm across these generations while preserving global shutter behaviour [S3]. The practical effect is that global shutter is no longer the resolution penalty it was a decade ago, which is closing the price gap that used to justify a rolling shutter on cost alone. Interface-side, GigE Vision, USB3 Vision, and CoaXPress 2.0 are the dominant transport standards for in-line cells, and trigger-over-cable and PTP time synchronisation are now table stakes for multi-camera global shutter rigs.
On the application side, the cobot risk assessment workflow is the right place to confirm that a robot-mounted vision cell is being specified as moving or stationary at the moment of capture, because ISO/TS 15066 power-and-force-limiting calculations assume a known cell state, and a rolling shutter skew error is the kind of geometric drift that quietly invalidates a cell layout that was signed off as stationary. The predictive maintenance decision guide is the natural follow-on once the cell is in production, because the strobe synchronisation, trigger cabling, and enclosure interlock are exactly the failure modes a CMMS work order should track.