Event-based (neuromorphic) cameras
Also known as Event cameras, Dynamic vision sensors, Neuromorphic vision sensors
Event-based cameras have pixels that report only when the brightness at that pixel changes, instead of recording full frames at a fixed rate. This gives very fast response, a wide range of brightness that a camera can handle, and low data output, and it needs new software to use well.
In plain language
A normal camera takes a series of still pictures, many times a second, and every picture contains every pixel whether or not anything changed. An event-based camera works differently. 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. A still scene produces almost no messages, while a fast-moving edge produces many.
Because each pixel reacts on its own, the camera can respond in microseconds and copes well with scenes that have very bright and very dark areas at once, such as a car leaving a dark tunnel. The output is a stream of timed events rather than images, so the software that reads it has to be designed for that stream. Processing methods are an active area of research.
Technical explanation
Event-based sensors, also called dynamic vision sensors, use a per-pixel change detector. Each pixel compares its log-intensity with a stored reference level and emits an event with an address, a timestamp and a polarity (brighter or darker) whenever the change crosses a threshold, then resets the reference. An early silicon version was reported by Lichtsteiner, Posch and Delbruck: a 128 by 128 asynchronous temporal contrast sensor in IEEE Journal of Solid-State Circuits in 2008, with a 120 dB dynamic range and 15 microsecond latency. Brandli and colleagues reported a 240 by 180 global shutter sensor in 2014, with 130 dB dynamic range and 3 microsecond latency.
The survey by Gallego and colleagues sets out the sensor principle, the vision tasks the sensors support (from feature tracking and optic flow to reconstruction and recognition) and the processing methods used, including learning-based approaches and spiking neural network hardware. The survey reports that event cameras offer temporal resolution in the order of microseconds and a dynamic range of 140 dB, compared with 60 dB for conventional cameras, while consuming little power and reducing motion blur. These are comparative figures from a review, and the abstract does not state test conditions for each.
Because the output is sparse and asynchronous, standard frame-based algorithms do not transfer directly. Deep learning methods need new representations of the event stream, and a 2023 survey and benchmark by Zheng and colleagues reviews these approaches. Commercial sensors now exist: a company developer page for the Sony and Prophesee IMX636 gives a 1280 by 720 array with 4.86 micrometre pixels, a dynamic range beyond 120 dB and a peak output of 1 giga-event per second, though this page does not give test conditions.
What the evidence says
Each statement is typed by the strength of its evidence and linked to the source that supports it.
Demonstrated Shown in a published experiment, test or shipping product.
Event-based cameras capture per-pixel brightness changes asynchronously, instead of recording full frames at a fixed rate.
Sources: [1] · AbstractA 128 by 128 asynchronous temporal contrast sensor reported in 2008 had a 120 dB dynamic range and 15 microsecond latency.
Sources: [2] · TitleA 240 by 180 global shutter event sensor reported in 2014 had a 130 dB dynamic range and 3 microsecond latency.
Sources: [3] · TitleGallego and colleagues report that event cameras offer temporal resolution in the order of microseconds and a dynamic range of 140 dB, compared with 60 dB for conventional cameras.
Sources: [1] · Abstract · medium confidenceA 2023 survey and benchmark by Zheng and colleagues reviews deep learning methods for event-based vision.
Sources: [4] · TitleThe Gallego survey covers event camera processing from feature tracking and optic flow through to reconstruction and recognition, and it reports that event cameras reduce motion blur.
Sources: [1] · Abstract · medium confidenceReported, unverified Stated by a company or team but not independently confirmed.
The IMX636 HD event-based sensor is documented with a 1280 by 720 pixel array, 4.86 micrometre pixels, a dynamic range beyond 120 dB and a peak output of 1 giga-event per second.
Sources: [5] · Specifications · medium confidenceLimitations and open problems
- Output is a stream of events rather than frames, so most existing computer vision software cannot use it directly and new algorithms and representations are needed (Gallego and colleagues, 2022; Zheng and colleagues, 2023).
- Event sensors respond to change, so a scene with little motion produces little information, and features that do not move may be missed unless the camera or the scene moves.
- The sources reviewed here report sensor figures from the makers and from research papers, but they do not give a full cost comparison with frame cameras or market volumes.
- Sensor figures such as dynamic range are often given without test conditions in company material, so they should be read as reported rather than independently verified.
Possible future uses
Ideas, not predictions. None of these has been demonstrated.
- Vision for drones and small robots that must react within milliseconds while running on small batteries.
- Monitoring of fast industrial processes, such as high-speed production lines, where frame cameras miss short events.
- Driver monitoring and night-time vision systems that cope with sudden changes in light.
Product concepts: Portable unit that warns drivers before they hit a parked car
Timeline
- 2008Lichtsteiner, Posch and Delbruck report a 128 by 128 asynchronous temporal contrast sensor with 120 dB dynamic range and 15 microsecond latency in IEEE Journal of Solid-State Circuits. Sources: [2]
- Oct 2014Brandli and colleagues report a 240 by 180 global shutter event sensor with 130 dB dynamic range and 3 microsecond latency. Sources: [3]
- Jan 2022The survey 'Event-based Vision: A Survey' by Gallego and colleagues appears in IEEE Transactions on Pattern Analysis and Machine Intelligence, volume 44, issue 1. Sources: [1]
- Feb 2023Zheng and colleagues post a survey and benchmark of deep learning methods for event-based vision as a preprint. Sources: [4]
Futures this could help bring about
Related breakthroughs
Glossary
- Event
- A short message from one pixel saying that its brightness has changed by more than a set amount, with a time stamp and a direction (brighter or darker).
- Dynamic range
- The ratio between the brightest and darkest light a sensor can record in one scene, often given in decibels (dB).
- Optic flow
- The apparent movement of objects and edges across the field of view, caused by motion of the camera or the scene.
- Spiking neural network
- A type of neural network that passes information as discrete spikes in time, which suits the event output of these sensors.
Sources
- peer reviewed paper
Event-based Vision: A Survey
Guillermo Gallego, Tobi Delbruck, Garrick Orchard et al.. IEEE Transactions on Pattern Analysis and Machine Intelligence, Jan 2022. DOI: 10.1109/TPAMI.2020.3008413. Accessed 21 Oct 2026. - peer reviewed paper
A 128×128 120 dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor
Patrick Lichtsteiner, Christoph Posch, Tobi Delbruck. IEEE Journal of Solid-State Circuits, 2008. DOI: 10.1109/JSSC.2007.914337. Accessed 21 Oct 2026. - peer reviewed paper
A 240 × 180 130 dB 3 µs Latency Global Shutter Spatiotemporal Vision Sensor
Christian Brandli, Raphael Berner, Minhao Yang et al.. IEEE Journal of Solid-State Circuits, Oct 2014. DOI: 10.1109/JSSC.2014.2342715. Accessed 21 Oct 2026. - preprint
Deep Learning for Event-based Vision: A Comprehensive Survey and Benchmarks
Xu Zheng, Yexin Liu, Yunfan Lu et al.. arXiv, 17 Feb 2023. Accessed 21 Oct 2026. - company announcement
IMX636 HD Sensor
Prophesee developer documentation, 21 Oct 2026. Accessed 21 Oct 2026.
Revision history
3 changes since this profile was created.