Menu

Event-Based Vision

Event-Based Vision: Cameras That Only Report Change

NotesNovember 19, 202611 min readSenova Systems Research

An event camera never takes a picture. Each pixel fires only when the brightness changes, in under 100 microseconds, across a million-to-one range of light. What that buys, what it costs, and when to pick it over frames.

First, a simple picture

A normal camera is a metronome with a lens. Every 33 milliseconds, thirty times a second, it takes a full picture of everything, whether anything changed or not. A quiet warehouse at 3 a.m. produces the same stack of pictures as a busy loading dock at noon. The camera does not know or care what moved. It reports everything, always.

An event camera is a room full of watchers. Each one watches a single tiny patch of the scene, all the time, and says nothing, until the patch changes. Then it speaks, once: "this spot got brighter, right now." A quiet warehouse produces almost no words at all. A fast forklift produces a burst, exactly where it is, exactly when it moved.

That is the whole idea. An event camera, also called a dynamic vision sensor (DVS) or an event-based vision sensor, does not take pictures. Each pixel works alone, watches the brightness at its own spot, and fires an event only when the brightness changes enough. The output is not a video. It is a stream of tiny messages, each one saying where, when, and whether the light went up or down.

This is the sensing half of the story from Article #1 (neuromorphic computing). There, the chip was the brain: neurons that stay silent until something happens. Here is the eye that feeds it. An event camera speaks in spikes, and a spiking chip understands spikes. Together they form a pipeline that does almost no work when nothing is happening, and reacts in microseconds when it is.

How one pixel works

Four steps, and the pixel never leaves its post:

  1. Watch. The pixel watches its own patch of the scene, continuously. There is no shutter and no clock telling it when to look.
  2. Compare. It compares the brightness now with the brightness a moment ago. The comparison is done on the logarithm of the light, a scale where each step means "ten times brighter," like the Richter scale for earthquakes. That log scale is what gives the sensor its enormous range of light (below).
  3. Fire, or stay silent. If the change passes a set line, the contrast threshold, 25% on the Sony/Prophesee IMX636, the pixel fires one event: its x and y position, a timestamp measured in microseconds, and a sign: brighter (+) or darker (−). If the change is below the line, the pixel says nothing at all.
  4. Reset and watch again. The pixel forgets the old brightness and takes the new one as its baseline. The next change is measured against the new value.

There is no frame. There is no exposure time. There is no moment when the whole sensor is read at once. The stream is asynchronous, each pixel reports whenever it has news, in the order the news happens. That single difference explains every advantage and every limit that follows.

How one event pixel works: no shutter, no frame, no clock

StepWhat the pixel does
1. WatchWatches its own small patch of the scene, all the time
2. CompareCompares the brightness now with a moment ago; fires above a 25% change (IMX636; early sensors 10 to 15%)
3. SpeakSends one event: x, y, time and brighter or darker, timed to the microsecond
4. SilenceNo change, no event: no data and no power

What it buys, with real numbers

Speed. The Sony/Prophesee IMX636, the current high-resolution event sensor, 1280×720 pixels, built on a modern stacked process, reports a change in under 100 microseconds at 1,000 lux (ordinary indoor light). A frame camera at 30 frames a second cannot report anything sooner than one frame: 33 milliseconds. The event camera's first news arrives roughly 330 times sooner. The first practical DVS, built by Lichtsteiner and Delbruck in 2008, already answered in 15 microseconds, the speed is not new; the resolution is.

No motion blur. A frame camera exposes each picture for some time, a millisecond, a few milliseconds, and anything moving during that window smears across the picture. An event camera has no exposure window: each event carries a microsecond timestamp. A fast object is a crisp trail of events, not a blur. One early DVS paper proved the point by tracking a disc spinning at 10,000 revolutions per second, a frame camera would have needed 100,000 frames a second to do the same.

Light range. Dynamic range is the range of brightness a sensor can handle, from the dimmest detail to the brightest, in one picture. It is measured in decibels (dB), a log scale, so every 20 dB means ten times wider. A good frame camera manages about 60 dB: a thousand-to-one range. The event camera survey by Gallego and colleagues puts event cameras at 140 dB, a ten-million-to-one range. The IMX636's own datasheet guarantees more than 86 dB from 5 to 100,000 lux, and more than 120 dB down to 0.08 lux, moonlight. A frame camera pointed at a doorway, bright sun outside and a dim room inside, must choose: expose for the room and blow out the door, or expose for the door and lose the room. An event camera sees both, because each pixel adapts to its own patch of light.

Power that follows activity. The IMX636 draws 5 milliwatts on standby and 205 milliwatts at maximum, the sensor's power scales with how much is happening. A bandwidth study of silicon retinas (MDPI, 2019) puts typical DVS cameras at 10–20 milliwatts, against hundreds of milliwatts for conventional frame sensors: about a ten-fold cut. Prophesee's smaller GenX320 sensor, 320×320 pixels, aimed at always-on vision, idles in the tens of microwatts, by the maker's figures. Power that follows activity is the same trick as the spiking chip in Article #1, and it is why the two belong together: a quiet scene costs almost nothing, end to end.

Data that follows activity. This one is easy to count, and the demo below does it step by step.

Event camera against frame camera (Sony/Prophesee IMX636 datasheet 2024; Gallego et al. survey)

PropertyEvent cameraFrame camera
Time to first newsUnder 100 µs at 1,000 lux33 ms (one frame at 30 fps)
Dynamic rangeOver 120 dBAbout 60 dB
Motion blurNone (µs timestamps)Yes, at speed
Data when nothing movesNear zeroFull frames, always
Sensor power5 mW standby, 205 mW maximumHundreds of mW

Where it wins today

  • Fast motion, close up. A drone dodging, a robot arm moving, a part flying down a conveyor. The microsecond timestamps are the win, the system knows where the thing is now, not where it was a frame ago.
  • Low light and hard light. Night driving, tunnels, dawn and dusk, a camera looking out of a dark warehouse into sunlight. A November 2026 paper (Feng & Zhou, arXiv) tested event-based object detection on construction sites under exactly these conditions: in low light, 70% of their test set, the best event detector scored 27.2 AP against 21.3 AP for RGB, and fusing both sensors scored best of all at 33.3 AP. (AP is average precision, the standard score for "how well did it find the things.")
  • Always-on sensing on a battery. A doorbell, a wildlife camera, a machine monitor. When nothing moves, the sensor sends almost nothing, and the spiking chip behind it (Article #1) spends almost nothing. Prophesee's GenX320 class of sensor is built for exactly this.
  • High-speed 3D. Prophesee's EVK3D evaluation kit pairs the IMX636 with a fast laser projector and reports depth maps at 500 Hz with under 1.5% error (the maker's figures). A frame camera's depth map tops out at its frame rate.

Where it doesn't win: yet

Plain honesty, because this is where the hype usually hides:

  • A still scene is invisible. If nothing changes, the sensor sends nothing. A parked car, a still face, a barcode sitting on a table, the event camera has no picture of them. You need a frame camera alongside, or a way to reconstruct a picture from the event stream, or a sensor that does both (the DAVIS family pairs an event sensor with a frame sensor on one chip). An event camera alone cannot answer "what is in this room right now?"
  • Colour is rare and young. Most event sensors are monochrome, they measure brightness, not colour. Colour event sensors exist, but they are newer and lower in resolution. If the job needs colour (a sorting line that sorts by colour, say), frames still win.
  • The tools are young. The algorithms, the labelling tools, the debugging views, all of it is thinner than the frame-camera world, where fifty years of tooling exists. A 2026 study on low-latency machine vision makes the point plainly: sensor-level speed does not automatically become application-level speed. The pixel answers in 100 microseconds, but the rest of the system, the processor, the bus, the algorithm, has to keep up, and on a small edge board it often doesn't. (That study still measured person detection at 92 milliseconds, gesture recognition at 20 milliseconds, and SLAM at 15.9 milliseconds on small ARM boards, fast, but not microsecond-fast.)
  • Busy scenes can cost as much as frames. An event camera's data rate grows with motion. A busy, high-contrast scene can produce events at a rate that matches, or exceeds, a frame camera's data rate. The sensor's maximum is 1.06 billion events a second on the IMX636. The win is on quiet and sparse scenes; the demo below shows the arithmetic both ways.

The decision test

Four questions. Three yeses, and event-based sensing is worth a serious look:

  1. Is the scene mostly quiet, with bursts of motion? A door, a conveyor, a road at night, yes. A crowd, a busy street in daylight, probably not.
  2. Is speed the wall, microseconds, not milliseconds? Dodging, tracking, reacting, yes. Inspecting a still part, no.
  3. Is the light hard, very dim, very bright, or both at once? Yes is the event camera's home ground.
  4. Can the job live without a still picture and without colour? If the answer is no, keep a frame camera in the system, often with the event camera, as the 2026 construction-site paper shows: the fusion of both scored best in every lighting condition.

DEMO

The setup is made up; the rates are the published ones above. One second, one camera, one warehouse at night. A forklift crosses the frame halfway through the second. Both cameras are 1280×720. The frame camera shoots 30 frames a second, one byte per pixel (black and white). Each event is about 8 bytes (Prophesee's EVT 2.1 format packs one event into 64 bits).

Step 1, The frame camera. One frame is 1280 × 720 = 921,600 bytes. At 30 frames a second, that is 27.6 megabytes every second, the same whether the warehouse is quiet or busy, because the camera reports everything, always. (Colour triples it to about 83 MB/s.)

Step 2, The event camera, quiet half-second. In the first half-second, almost nothing moves: a few pixels flicker as a light hums. A published bandwidth study measured a quiet outdoor sequence at 28,670 events a second. At 8 bytes each, that is about 0.23 megabytes a second, roughly 120 times less than the frame camera, for the same scene.

Step 3, The event camera, busy half-second. The forklift crosses. Suppose the burst reaches 1 million events a second, a busy scene, a lot of edges moving. That is 8 megabytes a second. Still about 3 times less than the frame camera, but the gap has narrowed a lot. At the sensor's maximum rate, the event camera would send more than the frame camera. The honest line: the event camera's win is proportional to how quiet the scene is.

Step 4, Read it honestly. The frame camera pays the same 27.6 MB/s for a scene with nothing in it and a scene with everything in it. The event camera pays for what happens, and nothing else. For an always-on sensor on a battery, that is the difference between a product that lasts a season and a product that lasts a week. For a busy scene, the right answer is often both cameras, and the 2026 construction-site paper backs that up with measurements, not opinion.


"But we already have cameras", and two more fair doubts

"Our frame cameras work fine." For most jobs, they do, and this article is not a call to rip them out. The question is narrower: is there one place where frames are failing, the fast part, the dark corner, the battery-powered node? One event camera, in one place, is a cheap experiment. The fusion result from the 2026 paper is the pattern to copy: keep the frame camera, add the event camera where it earns its keep.

"The tooling is too young for production." It is younger than frame-camera tooling, and the 2026 low-latency study says plainly that the rest of the system has to keep up. But the sensors are no longer lab parts: Sony and Prophesee ship the IMX636 in volume, evaluation kits exist, and the software suite is a real product. The honest test is a bench evaluation on your scene, not a debate about maturity in general.

"A still scene is invisible, that's a deal-breaker." It is a deal-breaker for a camera that must photograph a still room. It is not a deal-breaker for a sensor that must notice motion, measure speed, or watch for a change. Name the job first. If the job is "notice," the event camera is the right tool. If the job is "show me," keep the frame camera, or pair them.

Sources

  1. Sony Semiconductor Solutions / Prophesee, IMX636-AAMR-C product brief and datasheet (2024)
    What we used

    1280×720, 4.86 µm pixel, latency <100 µs at 1000 lux / <1000 µs at 5 lux, dynamic range >86 dB (5–100k lux) / >120 dB (0.08–100k lux), contrast threshold 25%, max event rate 1.06 Geps, standby 5 mW, max 205 mW; EVK4 dev kit <550 mW under maximum event conditions

  2. Gallego, Delbruck, Orchard et al., "Event-based Vision: A Survey"
    What we used

    ArXiv 1904.08405 (updated 2024), event cameras: ~140 dB dynamic range vs ~60 dB for standard cameras, microsecond temporal resolution, no motion blur, low power; applications in robotics, tracking, surveillance, recognition, gesture control

  3. Lichtsteiner, Posch & Delbruck, "A 128×128 120 dB 15 µs Latency Asynchronous Temporal Contrast Vision Sensor"
    What we used

    IEEE J. Solid-State Circuits (2008), the first practical DVS: 120 dB, 15 µs latency, 1 Meps

  4. Leñero-Bardallo, Serrano-Gotarredona & Linares-Barranco, IEEE J. Solid-State Circuits
    What we used

    128×128 DVS, 10% contrast sensitivity, latency under 3.6 µs, tracks objects at 10,000 revolutions per second

  5. MDPI Sensors (2019), "Bandwidth Modeling of Silicon Retinas for Next Generation Visual Sensor Networks"
    What we used

    DVS 10–20 mW vs hundreds of mW for frame sensors (~10× cut), ~100× speed, 120 dB vs 60 dB dynamic range; quiet outdoor sequence measured at 28.67 kilo-events per second

  6. MDPI Electronics (2026), "Low-Latency Machine Vision Based on a Neuromorphic Vision Sensor"
    What we used

    Sensor-level latency does not automatically become application-level latency; person detection 92 ms, gesture recognition 20 ms, SLAM 15.9 ms on small ARM boards; logarithmic photoreceptor gives >100 dB dynamic range

  7. Feng & Zhou (2026), "Event-based object detection under adverse lighting and motion"
    What we used

    ArXiv 2611.19746 (published Nov 17, 2026), low-light event detector 27.2 AP vs 21.3 AP for RGB; late fusion of both sensors best in every lighting condition (33.3 AP)

  8. Prophesee EVK3D (2026)
    What we used

    IMX636 + VCSEL projector: depth maps at 500 Hz, <1.5% RMSE (maker's figures)

  9. Prophesee GenX320 figures via secondary summary (KPSphere, 2026)
    What we used

    320×320, idles in the tens of microwatts, operates in low milliwatts; labelled as maker-reported figures