Global Shutter vs Rolling Shutter: Choosing a Camera for Robotics and Machine Vision

Global shutter vs rolling shutter is the difference between an image sensor that starts and stops the exposure of every pixel at the same instant and one that exposes and reads the frame one row at a time from top to bottom. With a global shutter each frame is a single snapshot in time; with a rolling shutter the bottom of the image is captured slightly later than the top.

For a still scene the two look the same. When the camera or the subject moves, the difference shows up as bent verticals, wobble and smeared edges on a rolling shutter, and that matters for robots, vehicles and inspection lines, where software has to measure what it sees rather than just display it.

This guide explains how each shutter reads the sensor, what the rolling shutter effect looks like and how to calculate it, where pseudo-global shutter and strobes fit, how noise and dynamic range compare, and which shutter suits robotics, SLAM and machine vision. It finishes with sensor examples, the Stereolabs ZED range as a worked case and a checklist.

The table sets the two side by side. Entries describe typical sensors of similar size and cost; individual sensors vary.

Neither shutter is better in general. The right one depends on how much the scene moves relative to the camera during a frame, and on whether the system measures the image or only looks at it.

Every image sensor collects light in its pixels and then reads the charge out row by row. A rolling shutter staggers the exposure to follow that readout, so each row starts and ends its exposure a little after the row above. The time it takes to read the whole frame, often a few milliseconds up to tens of milliseconds, is the window in which motion can distort the image.

A global shutter sensor adds storage in each pixel, so the whole array can stop exposing at once and hold its charge while the rows are read out. That extra circuitry is why global shutter pixels have tended to be larger and the sensors more expensive for the same resolution, although the gap has narrowed as sensor technology has improved.

Exposure time matters as well as shutter type. A short exposure reduces motion blur on either sensor, but on a rolling shutter it does not remove skew, because the rows are still read at different times. Bright, even lighting lets a global shutter camera use short exposures freely, which is one reason vision systems pair them with strobes or high-power LED lighting.

The rolling shutter effect is the distortion that appears whenever something moves noticeably during the frame readout. The common forms are worth recognizing before they appear in your data.

• Skew. Vertical lines lean when the camera pans or a vehicle drives past, because each row sees the scene a moment later.
• Wobble. Vibration on a vehicle or drone makes the whole image ripple, often called the jello effect.
• Partial exposure. A short flash or strobe lights only some rows, leaving bright and dark bands across the frame.
• Flicker banding. Pulsed LED lighting can print horizontal stripes across the image.
• Spinning objects. Fans, propellers and rollers appear bent or broken into curved segments.

To a person watching a video these are cosmetic. To a perception model or a measurement algorithm they are errors, because an edge that bends or moves between rows gives the wrong position or size.

How to calculate it. The skew between the top and bottom of the frame is the speed of the image across the sensor, in pixels per second, multiplied by the readout time. The table works three cases with a 1920-pixel-wide sensor and a 15 ms readout, a typical figure for a full-resolution rolling shutter sensor.

The skew is several times the blur in every case, which is why shortening the exposure does not cure a rolling shutter problem. A useful rule: if the skew works out at more than a pixel or two for the edges the software measures, start with a global shutter, or reduce the readout time by using a region of interest.

There is a middle way for systems that control their own lighting. In a pseudo-global, or global reset, mode, a rolling shutter sensor starts the exposure of every row at the same time, then reads the rows out in turn. If the scene is dark apart from a strobe, and the strobe fires only while every row is exposing, each row records the same brief flash. The strobe duration becomes the effective exposure, and the image is free of skew.

It works when three conditions hold: the exposure must be longer than the readout so all rows overlap, the strobe must fit inside that overlap, and ambient light must be low enough that it adds little to the image before and after the flash. That suits enclosed inspection tunnels and some barcode readers. It does not suit robots and vehicles working in daylight, where a true global shutter is the dependable choice.

Often slightly more, though the gap is closing. The storage element in each global shutter pixel takes space and adds a source of read noise, so for the same pixel size a global shutter sensor has traditionally had a little more noise and less dynamic range than a rolling shutter one. Light leaking into the storage node during readout, known as parasitic light sensitivity, can also add faint smear on bright objects; manufacturers quote it as shutter efficiency.

Back-illuminated designs and newer pixel structures have narrowed the difference, and most machine vision systems use enough light that it rarely decides the choice. The Stereolabs product data shows the typical trade-off: the 2.3MP global shutter ZED X One GS quotes 71.4 dB of dynamic range, while the 8.4MP rolling shutter ZED X One 4K quotes 85 dB with HDR. Where a scene has deep shade and bright sun together, that extra range can matter more than motion.

Choose a global shutter when the camera moves or the subject moves fast. Autonomous mobile robots and vehicle systems see the world sweep past while they drive and turn; drones vibrate; parts on a conveyor pass the camera at speed; robot arms swing the camera with the gripper. Stereo depth is a particular case, because the camera computes distance by matching the same point in two images, and that matching is most reliable when both images and every row in them were captured at the same instant. This is why global shutter sensors are common in depth cameras for robotics.

Visual odometry and visual-inertial SLAM add a further reason. Most algorithms, including the visual SLAM packages common in ROS, assume each frame was captured at one instant and fuse it with IMU data at that timestamp. On a rolling shutter the rows have different timestamps, so fast turns introduce errors that show up as drift unless the software models the readout. A global shutter camera with a hardware-synchronised IMU removes that source of error, which is why depth cameras for navigation pair the two.

Choose a rolling shutter when detail, dynamic range and low light matter more than motion. Reading labels and text, inspecting fine features on a part that stops under the camera, monitoring a scene from a fixed mount, or recording in mixed lighting all suit the higher resolution and HDR that rolling shutter sensors often bring.

Multi-camera systems add one more reason to prefer a global shutter. When several cameras around a vehicle or robot are hardware-synchronised, each frame from each camera represents the same instant, so software can stitch surround views or fuse depth from several viewpoints without correcting for row timing. Where cameras on separate computers must agree, IEEE 1588 PTP aligns their clocks. If a system mixes the two, keep measurement and fusion on the global shutter cameras and use the rolling shutter camera for detail and recording. Our explainers on sensor fusion and LiDAR vs camera cover how cameras combine with other sensors.

Sensor makers offer both types across the range. These are examples from published sensor families; check the current data sheet for each part, as resolutions and modes vary.

Sensor size and pixel size matter as much as the shutter. Larger pixels gather more light, which supports shorter exposures on a moving machine, and the lens has to resolve the pixel pitch. Our computer vision explainer covers how the image feeds the model.

The Stereolabs range covers both, which makes it a useful worked example. The ZED X, ZED X Mini and ZED X Nano stereo cameras use dual 2.3MP global shutter sensors with 3 µm pixels, as do the ZED X One GS and ZED X One S monocular cameras and the board-level ZED X One Core, all built for moving robots and vehicles. The USB ZED 2i and ZED Mini use 4MP rolling shutter sensors, and the ZED X One 4K uses an 8.4MP rolling shutter sensor with HDR for detail and difficult lighting.

A common pattern is to use global shutter GMSL2 cameras for navigation and depth on a moving machine, and add a high-resolution rolling shutter camera where the job is to read or record detail rather than measure motion. The GMSL link matters too: it carries hardware frame-sync to every camera, so several global shutter cameras expose at the same moment. Our guide to the GMSL camera interface explains how ZED X cameras connect to a Jetson, and our explainer on GMSL3 covers the bandwidth that 4K cameras at high frame rates need.

Work through four questions in order.

• Motion. How far does the scene move relative to the camera during one frame readout? If it is more than a pixel or two, start with a global shutter.
• Measurement. Does software measure positions, depth or sizes from the image? Measurement is less forgiving of distortion than viewing.
• Lighting. Will strobes, flashes or pulsed LEDs light the scene, or does the camera face bright sun and deep shade? Strobes favor global shutter; high contrast may favor an HDR sensor.
• Resolution and budget. If the scene is slow and detail matters, a rolling shutter gives more pixels for the cost.

We specify cameras together with the computer and interface that run them, and we configure, image and test the whole system as standard. The computer matters because several global shutter streams plus depth processing need GPU capacity: our rugged Jetson computers and ZED Box Mini are the usual hosts. Our guide to setting up a ZED X robotics stack on Jetson covers the software that turns those images into depth and odometry, and our robot computer guide covers the rest of the onboard compute.