Prophesee - Retinal Event-Based Neuromorphic Vision — Prophesee

Biological model: Human retina photoreceptor and ganglion cell arrays

Company: Prophesee

Human retinal architecture inspires event-based vision sensors generating 1000x less data than conventional cameras whilst capturing motion and edges.

The challenge

Conventional frame-based cameras generate enormous data volumes (gigabytes per minute), requiring substantial bandwidth and power for transmission and processing. Motion detection and edge recognition waste resources processing static regions. Event-based approaches could reduce data whilst improving temporal resolution.

Nature's strategy

Human retinal ganglion cells respond to light intensity changes, not absolute intensity, compressing visual information 1000-fold before transmission via optic nerve.

What was emulated

Event-driven encoding; change detection at pixel level; data compression via selective information transmission; temporal precision; low-latency motion detection.

The innovation

Silicon event camera where individual pixels emit timestamped events only when light intensity changes above threshold, mimicking retinal ganglion cell behaviour.

Full case study

The human retina does not transmit raw pixel values to the brain; instead, it performs dramatic data compression on-sensor. Photoreceptors detect light, but retinal ganglion cells—the retina's output neurons—only fire when light intensity changes rapidly. Stationary scenes generate almost zero neural activity; motion, flicker, and edges trigger spike bursts. This event-driven encoding reduces the optic nerve's data burden to approximately 1 Mbps despite observing a complex visual scene—a 1000-fold compression versus a conventional 1 gigapixel video stream. Prophesee engineered silicon chips replicating this retinal principle: thousands of pixels, each with its own photodiode and change-detection circuitry. When light intensity changes above a threshold, that pixel emits a time-stamped event; unchanged regions remain silent. These event cameras require minimal power, generate vast data reductions, and excel at motion detection and low-light vision where conventional cameras fail. Applications range from industrial robotics (tracking fast-moving objects at microsecond precision) to autonomous vehicles (robust edge detection and motion segmentation) to surveillance (detecting rare, meaningful events in hours of static footage). By mirroring retinal compression architecture, Prophesee achieved sensor performance impossible with conventional frame-based approaches.