Event-based cameras: sensors that only report change

An event-based camera is a vision sensor that does not record full pictures. Each pixel watches the light falling on it and sends a short message only when that light gets brighter or darker by a set amount. An ordinary camera records every pixel many times a second, whether or not anything has moved. Event-based sensors are also called dynamic vision sensors or neuromorphic vision sensors. They matter because, according to reviews and manufacturers, they can respond within microseconds and cope with scenes that have very bright and very dark areas at once. Whether they become useful beyond specialist work depends mostly on software, which has to be rebuilt to read a stream of changes rather than a sequence of frames.

How the sensor works

Each pixel compares the brightness it sees with a stored reference level. When the change crosses a set threshold, the pixel records its location, the time and a direction (brighter or darker), then resets its reference. A still scene produces almost no messages, while a fast edge crossing the view produces many. The output is a timed stream of events rather than images, so the software reading it needs a different design from the code used for ordinary video.

What has been demonstrated

Here, "demonstrated" means shown in a peer-reviewed paper, for the conditions the paper gives.

• A drone flown on event data. Paredes-Vallés and colleagues, in a 2024 paper in Science Robotics, trained a spiking neural network that takes raw event-camera data and outputs low-level control commands for a drone (Science Robotics). The vision part was trained on real event data and the control part in a drone simulator. The network ran on board on Intel's Loihi neuromorphic chip at 200 Hz, using 0.94 W when idle and an extra 7 to 12 mW when running the network. The abstract reports hovering, landing and sideways manoeuvres. It gives no details of wind, lighting, altitude, number of trials or flight length.

What is reported but not independently tested

These figures come from a review, a preprint, paper titles or manufacturers, so they should be read as reported rather than checked.

• Early silicon, from paper titles. Lichtsteiner, Posch and Delbruck's 2008 paper describes a 128 by 128 sensor, and its title gives 120 dB dynamic range and 15 microsecond latency (IEEE, via DOI). Brandli and colleagues' 2014 paper gives a 240 by 180 global shutter sensor with 130 dB dynamic range and 3 microsecond latency in its title (IEEE, via DOI). We have not read the paper texts, so these are the papers' stated specifications, and their test conditions are not given here.
• A review by Gallego and colleagues states that event cameras offer temporal resolution in the order of microseconds and a dynamic range of 140 dB, against 60 dB for conventional cameras. It also says they use little power and reduce motion blur. The abstract gives no test conditions for these comparisons (arXiv).
• A survey with benchmarks. A preprint by Zheng and colleagues, first posted in 2023, reviews deep learning methods for event data and reports benchmark experiments on image reconstruction, deblurring and object recognition (arXiv). The abstract does not name the datasets used.
• Prophesee's developer documentation for the IMX636 sensor, built with Sony, gives a 1280 by 720 pixel array and a dynamic range "beyond 120 dB". The page gives no measurement conditions for that figure. It reports typical power use of 50 to 70 mW and a peak below 100 mW, and a peak output of one billion events per second (Prophesee docs).
• Prophesee's evaluation kit page gives a dynamic range above 86 dB when measured between 5 and 100,000 lux, and above 120 dB when measured from a low-light cut-off of 0.08 lux. The page notes that 100,000 lux is a virtual upper limit that was not reached in experiment. Pixel latency is below 100 microseconds (Prophesee EVK4 page). The gap between the two dynamic range numbers shows why the measurement range matters.

There is some commercial use. Prophesee's recap of 2025 says IDS launched its uEye EVS industrial camera series, built on the IMX636. It also says CenturyArks launched a camera module on the same sensor, which Prophesee describes as the first commercial event-based vision camera module with a MIPI interface (Prophesee recap). On 11 March 2026, IDS and Prophesee signed a letter of intent "to jointly pursue the development of next-generation industrial vision systems". The announcement names no products, volumes or dates (Prophesee). The recap names no automotive customer.

Limits and open problems

• Software. Most computer vision software and training data assume full frames. Zheng and colleagues state that there is still a lack of taxonomies for deep learning techniques in event-based vision (arXiv), so the sources reviewed show no shared standard yet.
• Stillness. A scene that does not change produces little data. Features that do not move relative to the sensor can be missed unless the camera or the scene is moving.
• Resolution. The best-documented commercial sensor in this review has 1280 by 720 pixels. The sources reviewed do not give a like-for-like resolution comparison with frame cameras, so the trade-off is hard to judge.
• Cost and volume. The sources reviewed give no cost comparison with frame cameras and no market volumes.
• Evidence. Most performance figures come from manufacturers, or from papers whose test conditions are not given in the sources reviewed. Real-world trials are few in those sources, and the drone result is a single study whose abstract gives limited test detail.

Where it might go

The sources reviewed do not show the following uses yet. They are proposals, not results.

• Small drones and robots that must react within milliseconds while running on small batteries.
• Monitoring of fast industrial processes, where frame cameras can miss short events.
• Driver monitoring and night-time vision that cope with sudden changes in light.

The sources reviewed show no phone, mainstream robot or production car that uses an event sensor, so these uses remain plans rather than products.

What to watch next

• A named carmaker or supplier announcing a production vehicle that uses an event sensor.
• On-vehicle, real-time tests that report end-to-end latency and power use.
• An event sensor in a phone or consumer device, with the sensor maker named.
• A common data format and public benchmark adopted by several vendors.
• Whether the IDS and Prophesee partnership names products, volumes or dates.