Event Camera (EVS) White Paper
How an event camera (EVS) is built, how it sees motion, and how it differs from an ordinary RGB camera — explained with diagrams and video.
1. Summary
An ordinary camera (RGB camera) captures the whole picture at fixed intervals. An event camera (EVS, Event-based Vision Sensor) sends a signal only from pixels whose brightness changed, at the moment they changed. An EVS is therefore less a picture-taking device than a sensor that outputs motion directly.
| RGB camera | Event camera (EVS) | |
|---|---|---|
| What it outputs | Every pixel's brightness at fixed intervals (fps) = the state of the scene | (x, y, polarity, time) of pixels that changed = the change in the scene |
| Time resolution | 1/fps (33 ms at 30 fps), blurred during exposure | µs time stamps, equivalent to over 10,000 fps (spec) |
| Dynamic range | about 60–70 dB | over 120 dB (spec) |
| Data volume | Constant, whatever the scene | Proportional to motion, almost zero for a still scene |
| Good at | Colour, texture, recognising still objects | Fast motion, speed, backlight and low light, low latency |
| Not good at | Moments between frames, high-speed motion, strong backlight | Still objects, colour and texture |
In one line: an EVS sees "when, where and how fast" something moved; an RGB camera sees "what" it is. The two sensors complement each other more than they replace each other.
2. What a frame camera misses
Like film, a frame camera shows motion by stringing still pictures together quickly. This has three built-in limits.
- It cannot see between frames. A 30 fps camera takes one picture every 33 ms and records nothing in between. A fast flash, collision or fall can be over between two pictures.
- It blurs during the exposure. If an object moves while a picture is being taken, it smears. A shorter exposure makes the picture darker and noisier.
- It resends what has not changed. Even when most of the scene is still, every pixel is sent and processed in every frame, and finding motion means filtering the background out of all that data again.
On top of that, one exposure holds only a narrow range of brightness, so in strong backlight the bright side turns white and the dark side black.
3. How an event camera works
Each EVS pixel keeps watching the brightness at its position as a logarithmic (log) value. When the difference between the current value and the value at its last event (the reference) exceeds a set threshold C, it sends one event at that moment.
- C or more brighter than the reference → ON event; C or more darker → OFF event.
- After an event, the reference becomes the current value and the pixel waits for the next change.
- If the brightness does not change, nothing is sent.
- Every pixel does this on its own, all at the same time (asynchronously) — it does not wait for other pixels or a frame clock.
What one event contains
| Field | Meaning |
|---|---|
| x, y | Pixel position |
| Polarity | Brighter (ON, 1) or darker (OFF, 0) |
| Timestamp | In µs, stamped on each event |
Events are a list of "which pixel changed when, and which way", so to show them to people you collect the events over a period and draw them (accumulated frame, time surface — chapter 7).
Why log brightness
A difference of log values is a ratio of brightness. So the edge of the same object shows the same size of change in deep shade and in sunlight 400 times brighter. This is the basis of the wide dynamic range of an EVS (over 120 dB).
4. EVS sensor structure
Behind the light-sensitive part, each EVS pixel holds a circuit that decides whether something changed. Where a frame-camera pixel "collects light and waits to be read", an EVS pixel "decides for itself and reports only what changed".
| Stage | What it does |
|---|---|
| Photodiode | Turns incoming light into current |
| Log conversion (photoreceptor) | Turns current into a log-brightness voltage — handles everything from near darkness to direct sunlight at once |
| Change detector | Keeps computing the difference between the current value and the value stored at the last event |
| Comparators (ON / OFF) | Sends an ON event above +C or an OFF event below −C, then resets the reference |
| Asynchronous readout | Reads only pixels that fired, in order, and sends them with position, polarity and time |
Stacked sensor and on-chip processing
The Sony IMX636 used in ROVI-EVS is a stacked sensor: a pixel chip that receives light is bonded onto a logic chip that decides and processes events. Moving the decision circuits to the lower chip keeps the pixels small (4.86 µm) while reaching 1280 × 720 resolution. The logic chip also has functions that tidy up the events.
| On-chip function | What it does |
|---|---|
| Event rate control (ERC) | Caps events per second so a sudden flood cannot overrun transfer or processing |
| Anti-flicker (AFK) | Filters events caused by light that flickers at a fixed frequency, such as fluorescent or LED lighting |
| Noise filter (STC) | Filters isolated events not connected to their neighbours (noise) |
5. Frames and events — what is different
Draw a black dot circling a bright disc five times a second in x · y · time space and the difference is plain.
| RGB camera | Event camera (EVS) | |
|---|---|---|
| How it works | All pixels expose at the same time → read out together | Each pixel separately, the moment a change happens |
| Time resolution | 1/fps — 33 ms at 30 fps | µs time stamps, equivalent to over 10,000 fps (spec) |
| Latency | Exposure + waiting for frame transfer | Pixel latency 100 µs or less @ 1000 lx (spec) |
| Motion blur | Occurs during exposure | None — there is no exposure |
| Dynamic range | about 60–70 dB (HDR needs several exposures) | over 120 dB (spec) — log response |
| Data volume | 1280 × 720 × 60 fps = about 55 million pixel values a second, whatever the scene | about 0.33 million events/s in a near-still room, about 2 million events/s with people walking (measured) |
| Still scene | Visible as is | No events → looks empty |
| Colour and texture | Yes | No (brightness change only) |
| Sensor settings | Exposure, gain, white balance (mature automation) | Sensitivity and on-chip filters set to suit the scene |
6. Motion blur and dynamic range
Picture fan blades turning 20 times a second, with the right half in sunlight 400 times brighter than the shade.
- Motion blur: a shorter RGB exposure reduces blur but darkens the picture. An EVS has no exposure, so even fast objects keep their edges.
- Dynamic range: expose an RGB camera for the shade and the sunlit side saturates; expose for the sun and the shade goes black. An EVS sees changes in the ratio of brightness, so it sees both at once.
- Background: the still wall and hub do not appear. Only moving things remain, so no background removal is needed.
7. Real recordings
All of the following were recorded by FLOYD with a Triton2 EVS (6.5 mm lens).
30 fps and 10,000 fps
Drawing the events at 0.1 ms intervals lets you watch them like 10,000 fps video. Here it is next to the same scene as a 30 fps camera would capture it (33 ms exposure).
Two ways to look at events
| View | How it is drawn | Good for |
|---|---|---|
| Accumulated frame (events view) | Events over a short time as dots, coloured by polarity | Seeing edge shapes |
| Time surface | Each pixel's last event time as brightness — the more recent, the brighter | Seeing direction and flow at a glance |
Tracking moving targets
In an EVS, "where events gather is where something moved", so moving targets can be found directly, without a background model or an AI recognition model. ROVI-EVS gives each target an ID and works out its speed and trail — updates at 100 Hz, tracking p95 3 ms (measured).
8. Strengths and limits
Strengths
| Strength | Why | Figures |
|---|---|---|
| Motion = signal | A still background is not output → no background removal needed | — |
| High time resolution, low latency | Each pixel outputs the moment it changes | Equivalent to over 10,000 fps · pixel latency 100 µs or less (spec) |
| No motion blur | There is no exposure time | — |
| Wide dynamic range | Log response — sees ratios of brightness | Over 120 dB · down to 0.08 lx (spec) |
| Less data, less computing | Only changed pixels are output, almost zero when still | About 0.33 million events/s in a still room (measured) |
| Privacy | No colour or texture in the output, so faces are hard to recognise | — |
Limits
| Limit | Details | What to do |
|---|---|---|
| Still objects are invisible | When both camera and target stop, there are no events | Tracking software remembers briefly stopped targets · use RGB to confirm a still state |
| No colour or texture | It knows "something is moving" but finds it hard to tell people, vehicles and objects apart | Pair with an RGB camera when the type matters |
| Low contrast | Changes smaller than the threshold C produce no events — targets close to the background colour, long distances | Adjust lens, lighting and mounting distance |
| Event floods | Flickering lights, camera shake, or lights switching on and off make every pixel respond | On-chip anti-flicker and event rate control, fix the camera firmly |
| Resolution and price | 0.92 MP, 1/2.5" sensor — costs more than a comparable RGB camera | Use where motion and speed are what matter |
| Tools and experience | Fewer AI models, public data and standard tools than RGB; per-scene setting experience needed | FLOYD processing software and set-up support |
9. Where to use it
Selection guide
◎ best · ○ possible · △ limited · ✕ unsuitable
| Task | EVS | RGB | Both | Why |
|---|---|---|---|---|
| People counting, paths, passing speed | ◎ | ○ | ◎ | Sees only motion, no background handling |
| High-speed falling objects, rotating parts, vibration | ◎ | △ | ○ | µs time resolution, no blur (RGB needs a high-speed camera) |
| Backlit entrances, tunnels, night | ◎ | △ | ◎ | Over 120 dB dynamic range |
| Long-term monitoring with limited bandwidth or power | ◎ | △ | ○ | Almost no data when nothing moves |
| Privacy-sensitive spaces | ◎ | △ | △ | No faces or texture in the output |
| Telling vehicle and person types apart | △ | ◎ | ◎ | Type from RGB, motion and speed from EVS |
| Still objects, stock checks | ✕ | ◎ | ○ | An EVS cannot see what has stopped |
| Colour, text, appearance inspection | ✕ | ◎ | △ | Colour and texture are RGB territory |
Using both: the EVS reports first when and where something moved and how fast, and the RGB camera checks only that area at that moment. Each sensor covers the other's weak points, and the RGB side needs less computing.
Application examples
Application areas from the camera supplier's (LUCID) material. Figures are from the vendor material and depend on conditions.
| Area | Examples |
|---|---|
| Machines and processes | Non-contact vibration frequency measurement (1 Hz to kHz, pixel level), cable and yarn speed and slip monitoring, high-speed counting (over 1,000 objects a second), high-speed 3D with a line laser (over 10 kHz) |
| Safety and spaces | People counting and tracking, fall detection, safety-zone intrusion, gesture recognition |
| Traffic | Traffic counts and vehicle tracking, scenes with large brightness differences such as tunnel entrances |
| Drones and robots | Tracking fast flying objects, obstacle avoidance, LED marker positioning |
| People and gaze | kHz-class eye tracking, driver monitoring (low light) |
FLOYD is focusing first on tracking moving people and vehicles (the ROVI-EVS evaluation kit), high-speed counting of falling objects (pharmaceuticals), and tracking moving targets in backlight and low light.
10. Triton2 EVS and ROVI-EVS
LUCID Triton2 EVS
| Item | Details (spec) |
|---|---|
| Models | TRT009S-EC (Sony IMX636, 1280 × 720, 0.9 MP) · TRT003S-EC (Sony IMX637, 640 × 512, 0.3 MP) |
| Pixel | 4.86 µm · 1/2.5" (IMX636) |
| Time resolution | Equivalent to over 10,000 fps |
| Dynamic range | Over 120 dB |
| Low-light cut-off | 0.08 lx |
| Interface | M12 2.5GigE · PoE |
| Protection | IP67 (dust and water) |
| Lens | C-mount |
ROVI-EVS
ROVI-EVS is an evaluation kit combining the Triton2 EVS (TRT009S-EC) with FLOYD processing software. Using only a CPU, without a GPU, it finds moving targets, shows ID, box, speed and trail in real time and saves them as CSV, and lets you replay fast moments in 10,000 fps super slow motion.
| Item | Details |
|---|---|
| Update rate | 100 Hz (10 ms) (measured) |
| Tracking processing time | p95 3 ms (measured) |
| Live processing limit | 9.8 Mev/s, no loss (measured) |
| Super slow motion | 0.1 ms intervals, equivalent to 10,000 fps |
Set-up and use are in the ROVI-EVS user manual; send application review requests through floyd.kr contact.
11. Glossary
| Term | Meaning |
|---|---|
| EVS, event camera | Event-based Vision Sensor. Outputs only pixels whose brightness changed, at the moment they changed. Also called DVS (Dynamic Vision Sensor) or neuromorphic camera |
| Event | One brightness change at one pixel — (x, y, polarity, time) |
| Polarity | Brighter (ON) / darker (OFF) |
| Threshold (C) | The smallest log-brightness change that produces an event — smaller is more sensitive but noisier |
| Event rate | Events per second. ev/s, Mev/s (millions per second) |
| Accumulated frame | Events over a short time drawn together in one picture |
| Time surface | A picture showing each pixel's last event time as brightness |
| Dynamic range | Ratio between the darkest and brightest parts visible in one scene at once. 20 dB = 10×, 120 dB = 1,000,000× |
| Equivalent fps | Frames per second you get by drawing events at that interval. Equivalent to 10,000 fps = 0.1 ms intervals |
| ERC, AFK, STC | On-chip event rate control, anti-flicker and noise filter (chapter 4) |