What Is a Low Latency Camera and Why It Matters
An operator turns the wheel on a teleoperated vehicle. The steering responds immediately. The video showing the result arrives 180 ms later. By the time the operator sees the vehicle turning, they have already over-corrected, and the loop starts oscillating.
Nothing in that system was broken. The camera streamed at 60fps, the network was fine, and the encoder worked as designed. The delay accumulated across five stages that nobody added up.
This post is about that failure mode. A low-latency camera is not simply a fast camera, and frame rate does not tell you what you need to know. What follows is how to build a latency budget for low-latency camera streaming, where the milliseconds actually go, and which architecture removes them.
If you are specifying a real-time streaming camera for anything that controls a machine, end-to-end latency is the number to interrogate before resolution, frame rate, or sensor.
What a Low Latency Camera Delivers That a Fast Camera Does Not
Answer first: a low-latency camera is one where the delay between light hitting the sensor and the frame being usable by your application is short enough that your control loop stays stable.
Your application sets that threshold, not the camera. A surveillance recorder tolerates 200 ms. A teleoperated vehicle needs under 50 ms. A visual servoing loop on a robot arm may need under 20 ms.
So “low latency” is only meaningful against a number you have defined. Asking what is low latency camera without stating your loop rate makes the answer is unanswerable.
The measurement that matters is glass-to-glass latency, meaning the full path from photons at the lens to pixels on your display or in your algorithm’s buffer. Anything narrower hides the stages that usually dominate.
Camera latency and video latency get used loosely, and vendors quote whichever number flatters the product. When a datasheet gives a figure without saying which stages it covers, treat it as a sensor-and-interface number and assume your end-to-end latency will be higher. That is the practical answer to what is a low-latency camera: not a category, but a measured number against a stated requirement, and video latency claims mean little without one.
What Causes Camera Latency? The Five-Stage Budget
Latency accumulates in five stages, and in most systems the camera is not the biggest contributor. Adding them up is the only way to know which one to attack.
Stage | What it is | Typical driver |
Exposure | Integration time for one frame | Light level, gain |
Sensor readout time | Clocking pixels off the sensor | Resolution, lane count, shutter type |
ISP processing | Demosaicing, white balance, noise reduction | Pipeline depth |
Encoding | Compressing the frame | Codec choice, GOP structure |
Transport and host | Interface, protocol stack, buffering | USB, MIPI, GigE, WiFi, driver depth |
Exposure and Sensor Readout Time
Exposure is the floor. A 20 ms exposure means no frame can possibly arrive sooner than 20 ms after the event, which is why low light is a latency problem as well as an image quality one.
Sensor readout time follows, and it scales with how many pixels you are moving and how many lanes you have to move them on. A 20MP frame off four MIPI lanes reads out faster than the same frame off two.
Shutter type changes the shape of this. Global shutter latency is uniform across the frame because every row is exposed at once. Rolling shutter reads row by row, so the bottom of the frame is captured later than the top, which is a distortion problem and a timing skew at the same time. Global shutter latency is therefore easier to reason about in a control loop, because every pixel shares one timestamp.
ISP Processing Depth
Every stage in the image signal processor holds the frame for a period. Demosaicing, noise reduction, and HDR fusion each add pipeline depth.
This is why raw output can be a latency strategy. Skipping the ISP entirely and sending Bayer data to a host that processes it removes those stages from the camera path, though your host then has to do that work.
Encoding: H.264 Latency vs MJPEG Latency
Compression is often the largest single contributor and the most overlooked. H.264 latency depends heavily on GOP structure, because a frame that references future frames cannot be emitted until those frames exist.
MJPEG latency is lower because every frame is independent, at the cost of a much larger stream. Uncompressed video streaming removes the encoder from the path completely, which is the lowest latency option and the highest bandwidth one.
For real-time video streaming, the ordering is usually uncompressed first, then MJPEG latency, then H.264 latency, and the gap between the ends of that list is often larger than every other stage combined.
Transport and Host Buffering: Where the Stacks Live
The interface adds delay in two ways: the time to move the bytes, and the software layers they pass through on arrival.
This is where architectures diverge sharply. A USB stack, a network stack, and a direct sensor connection are not comparable, and the difference is usually larger than anything happening inside the camera.
Glass-to-Glass Latency: How to Measure It Properly
Point the camera at a running millisecond timer, put the live feed on a screen beside it, and photograph both at once. The difference between the two numbers is your glass-to-glass latency.
It sounds crude. It is also the only method that captures every stage, which is exactly why it is the standard technique.
Measure at your real settings. Latency at 1080p with short exposure tells you nothing about the same camera at full resolution in dim light, because exposure and readout both change.
Take many samples, not one. The variation between them is jitter in video streaming, and jitter often matters more than the average. A system with 40 ms average latency and 30 ms of jitter is harder to control than one with a steady 60 ms, because a control loop can compensate for a constant delay but not for an unpredictable one.
Record where you measured to. Latency to a display includes the display’s own processing. Latency to your algorithm’s buffer does not, and for a machine vision system that second number is the one that matters.
Frame Rate vs Latency: Why a Fast Camera Can Still Be Slow
This confusion costs more projects than any other item on this page, so it deserves its own section.
Frame rate tells you how often a new frame starts. Latency tells you how long each frame takes to arrive. They are independent, and a camera can be excellent at one and poor at the other.
A 120fps camera produces a fresh frame every 8.3 ms. If the pipeline behind it buffers three frames before delivery, your latency is still 25 ms or worse regardless of that frame rate.
The reverse also holds. A 30fps uncompressed stream over a direct connection can deliver lower latency than a 120fps compressed stream over a network, because the compression and the network stack cost more than the frame interval saves.
The practical rule for frame rate vs latency: frame rate sets your temporal resolution, latency sets your responsiveness. If you are building a real-time streaming camera system for control rather than for recording, latency is the specification to interrogate first and frame rate second.
Interface Choice: Low Latency MIPI, USB and GigE Cameras
The interface decides how many software layers sit between the sensor and your application, and each layer buffers data.
A low-latency MIPI camera has the shortest path. MIPI CSI-2 connects the sensor directly to the processor with no protocol stack in between. There is no enumeration step, no host driver layer of any depth, and no network. The cost is cable length, typically well under a metre.
A low-latency USB camera trades some of that for convenience. UVC camera latency is respectable and needs no driver development, which is why USB dominates prototypes and industrial PCs. The USB stack does add buffering, and the amount varies with host and kernel.
A low-latency GigE camera buys distance. GigE runs tens of metres on standard cabling, and the network stack adds delay and, more importantly, variability. On a quiet dedicated link this is manageable; on a shared network it is not.
Streaming protocols add their own tax. RTSP latency comes largely from the buffering that players insert to smooth playback, and that buffer is frequently larger than everything upstream of it combined. If you control both ends, shrink it.
None of these is universally correct. A low-latency MIPI camera wins on path length, a low-latency USB camera wins on integration time, and a low-latency GigE camera wins on distance. Match the interface to where your compute sits, then measure, because the theoretical ordering is often beaten in practice by a badly configured buffer.
Applications: Robotics, Drones, Teleoperation and Machine Vision
Some applications degrade gracefully as latency rises. Others stop working.
Teleoperation. A low-latency camera for teleoperation is a safety requirement rather than a comfort one. Human operators begin over-correcting somewhere around 100 to 150 ms of visual delay, which produces exactly the oscillation described at the top of this post.
Robotics and visual servoing. A low-latency camera for robotics feeds a control loop running at a fixed rate. If your frame arrives after the loop has already run, the controller acts on a stale position, and the error compounds every cycle.
Drones and aerial platforms. A low-latency camera for drones matters twice over: the pilot’s view and any onboard obstacle avoidance both depend on it. Airspeed converts every millisecond directly into metres of travel.
Machine vision and inspection. A low-latency camera for machine vision is about the reject gate. The part keeps moving while the frame is in flight, so a slow pipeline means the decision arrives after the part has passed the actuator.
Recording and analytics. This is the honest exception. If nobody acts on the frame in real time, latency is close to irrelevant, and you should spend your budget on resolution or dynamic range instead.
Read across those five, and the pattern is clear. A low-latency camera for robotics and a low-latency camera for machine vision are constrained by loop rate and actuator position, while a low-latency camera for drones and a low-latency camera for teleoperation are constrained by how fast a human or an autopilot can correct.
How to Reduce Camera Latency in a Real Design
Work the budget in order of size. The answer to how to reduce camera latency is rarely “buy a faster sensor”, and knowing how to reduce camera latency starts with knowing which stage currently dominates your latency budget.
Remove the encoder if you can afford the bandwidth. Uncompressed video streaming eliminates the largest and most variable stage in most pipelines. Where you cannot, choose a codec configuration without forward-referencing frames.
Shrink the receive buffer. Player and framework buffers are set for smooth playback, not for control. This is frequently a one-line change that saves more milliseconds than a hardware swap.
Shorten the exposure, if light allows. Exposure is a hard floor on latency. More light, a faster lens, or a more sensitive sensor buys you exposure time you can then spend on responsiveness.
Cut resolution to what the task needs. Fewer pixels read out faster and transport faster. This is the cheapest latency reduction available, and the one most often refused for the wrong reasons.
Move the ISP off the critical path. Raw output sends the frame without in-camera processing, which is why several teams choose Bayer output specifically for low-latency camera streaming rather than for image control.
Shorten the software path. Every abstraction layer between driver and application holds the frame briefly. Measure the difference between what the interface delivers and what your code receives, because that gap is often larger than expected.
The Vadzo Bolt Low Latency Camera Lineup
Vadzo’s Bolt series is the MIPI CSI-2 platform, connecting the sensor directly to the processor without a USB stack, host driver layer, or network in the path. That architecture is why the family is positioned for low-latency video streaming and for low-latency camera streaming into an embedded host.
Two tiers, and the distinction is worth respecting.
Modules with an explicitly stated sub-10ms latency figure
Module | Sensor | Resolution | Notes |
Onsemi AR2020 | 20MP colour, 5120 × 3840 | 4-lane MIPI CSI-2, uncompressed output, rolling shutter | |
Onsemi AR2020 | 20MP colour, VCM autofocus | 4-lane MIPI CSI-2 | |
Onsemi AR2020 | 20MP Bayer RAW | Host performs all demosaicing |
The Bolt-2020CRS product page states that four-lane MIPI CSI-2 outputs uncompressed data with sub-10ms latency. Both parts of that sentence matter: uncompressed removes the encoder stage, and four lanes shorten the readout.
Modules positioned for low latency by architecture
Module | Sensor | Resolution | Shutter |
Onsemi AR2020 | 20MP monochrome | Rolling | |
Onsemi AR0821 | 8.3MP, 4K HDR | Rolling | |
Onsemi AR0830 | 8MP, 3848 × 2168, HDR | Rolling | |
Onsemi AR0544 | 5MP | Rolling | |
Onsemi AR0235 | 2.3MP colour | Global | |
Onsemi AR0234 | 2MP monochrome | Global |
These are described by mechanism rather than by a number. The Bolt-830CRS brief puts it precisely: deterministic low latency, with no USB enumeration delay, no host driver layer overhead, and no network stack. For real-time video streaming, that removal is the whole argument. That sentence names the three sources of delay MIPI removes.
One distinction that trips people up. Low latency and global shutter are different properties. The sub-10ms modules above are rolling shutter, so a fast frame can still contain a distorted moving object. If your scene moves during exposure, you need both, and the Bolt-235CGS and Bolt-234MGS are the global shutter options in the family.
Read the tables against your budget. If your bottleneck is the encoder, a raw or uncompressed MIPI module removes it. If your bottleneck is readout, drop resolution or use all four lanes. If your bottleneck is the host, no module fixes it, and that is worth knowing before you buy one.
Frequently Asked Questions
What is a low latency camera?
A low-latency camera is one where the delay from light hitting the sensor to the frame being usable is short enough for your application’s control loop, and video latency claims should always be read against that requirement. Asking what is a low-latency camera without stating a target is unanswerable, because a surveillance recorder tolerates 200 ms while a robot arm may need under 20 ms. The figure that matters is glass-to-glass latency, covering exposure, sensor readout time, ISP processing, encoding, and transport together. Vendor figures often cover only part of that path, so confirm what your end-to-end latency budget actually includes. Vadzo’s Bolt MIPI CSI-2 series is built around this, with the Bolt-2020CRS product page stating sub-10ms latency on four-lane uncompressed output.
What causes camera latency?
Camera latency accumulates across five stages, and the camera is rarely the largest contributor. Exposure sets a hard floor, sensor readout time scales with resolution and lane count, the ISP holds the frame through each processing step, the encoder can dominate everything depending on codec configuration, and the transport plus host buffering adds the rest. Understanding what causes camera latency means measuring each stage rather than guessing, and it is usually where H.264 latency or host buffering hides. Knowing how to reduce camera latency follows directly from finding the dominant stage. Vadzo’s Bolt series removes several of these by design, using uncompressed video streaming over a direct MIPI CSI-2 connection with no USB or network stack, which is what makes low-latency video streaming achievable without touching the encoder settings.
Is a high frame rate camera the same as a low latency camera?
No, and treating them as equivalent is the most common mistake in this area. Frame rate tells you how often a new frame begins; latency tells you how long each one takes to reach your application. A 120fps camera producing a frame every 8.3 ms still delivers 25 ms of latency if the pipeline buffers three frames. Conversely, a 30fps uncompressed stream over a direct link can beat a 120fps compressed stream over a network, because MJPEG latency and inter-frame compression behave very differently under real-time video streaming conditions. For frame rate vs latency, specify latency first when the system controls something, whether that is a low-latency camera for robotics on a production cell or a low-latency camera for drones in flight.
Which interface gives the lowest camera latency?
MIPI CSI-2 generally provides the shortest path because the sensor connects directly to the processor with no enumeration step, no host driver layer, and no network stack. A low-latency USB camera adds a stack but requires no driver work, and UVC camera latency is acceptable for many applications. A low-latency GigE camera trades cable length for network variability, and RTSP latency is often dominated by player-side buffering rather than the network itself. A low-latency MIPI camera avoids all three layers. Cable length usually decides the answer before latency does. Vadzo’s Bolt series covers the MIPI case across sensors from 2MP to 20MP, which suits a low-latency camera for teleoperation or machine vision equally.
How do you measure camera latency?
Point the camera at a running millisecond timer, display the live feed on a screen next to it, and photograph both together. The difference between the two visible numbers is your glass-to-glass latency, and this method is used because it captures every stage rather than a convenient subset. Measure at the resolution, exposure, and settings you will actually deploy, and take many samples so you can see jitter in video streaming as well as the average. Repeat it with the encoder disabled to see how much of your latency budget compression is taking, and note that global shutter latency is easier to interpret because every row shares one timestamp. Vadzo’s engineering team can advise on measurement setup when you evaluate a Bolt module for low-latency camera streaming or as a real-time streaming camera in your own pipeline.
