Event Camera Fundamentals

Our cameras are based on the AlpsenTek HVS (Hybrid Vision) chip: the same pixel array simultaneously outputs an APS image stream and an EVS event stream. For the fusion architecture see HVS Hybrid Vision; for terminology see Glossary; for data formats see Data Formats.

Scene changes are recorded as an asynchronous event stream. Unlike a conventional frame-based image sensor, every pixel works independently and fires an event only when it detects a brightness change (e.g. object motion, a sudden light change).

This mechanism resembles how neurons in the human retina work: only dynamic information is transmitted, static background is ignored.

Event-camera principle illustration

Event-camera retina-like mechanism illustration

An event camera is composed of thousands of independently operating pixel units. Each pixel, with microsecond-level response speed, independently monitors brightness changes within its field of view. As soon as a brightness change exceeds a set threshold, that pixel immediately generates an "event".

Each event contains:

x, y: pixel position
t: timestamp (usually microsecond-level)
p: polarity (indicates whether brightness increased or decreased)

Event = (x, y, t, polarity)

Unlike the synchronous readout of frame images, events are generated asynchronously, sparsely, and continuously. This data format moves us from the "take a photo" paradigm into a new era of "perceiving change".

For example, when a golfer swings, the sensor captures only the trajectory of the ball and club, not the stationary sky or grass.

Pixel-level working principle

To understand where the above properties come from, look inside a single pixel:

Independent pixel operation: each pixel has its own circuit and monitors only the brightness at its own location, without depending on other pixels and without waiting for a global exposure — this is the physical source of "asynchrony".
Logarithmic brightness response: the pixel measures the logarithmic change of brightness (not linear brightness). The same proportional brightness change (e.g. doubling or halving) produces the same response magnitude regardless of whether the background is bright or dim — this is the root of the event camera's high dynamic range.
Threshold-crossing trigger: when the logarithmic brightness change accumulates past a set threshold, the pixel fires an event immediately, without waiting for other pixels or an "exposure period" — this is the source of microsecond-level response and zero motion blur.
Polarity (ON/OFF): brightening past the threshold → ON event (p=1); darkening past the threshold → OFF event (p=0). Pixels that stay unchanged emit no events — this is the source of sparsity, low power, and low bandwidth.
EPS (Events Per Second, event rate): the number of events generated per unit time, measuring how dynamically busy the scene is. Intense motion or jitter yields high EPS; a static scene has near-zero EPS.

One-line essential difference: a frame camera = "timed full-frame sampling" (fixed frame rate, all pixels synchronous); an event camera = "per-pixel asynchronous change-driven". Microsecond time resolution, low latency, sparsity, low power, high dynamic range — all stem from this fundamental difference.

Technical characteristics

Asynchrony and real-time performance: event data is output as a continuous stream with sub-millisecond response, avoiding the motion blur of conventional frame sensors. Each pixel triggers independently, so data generation is not bound by a fixed frame rate — suited to high-speed scenes (e.g. fast moving object tracking).
Data sparsity and low power: because only dynamic information is recorded, the event data volume is 1/10 to 1/1000 that of a conventional image sensor, significantly reducing compute demand and power consumption.
Strong environmental adaptability: the event sensor works stably under extreme lighting (low light or bright light), automatically adapting to brightness changes through independent pixel triggering, avoiding the under-/over-exposure problems of conventional sensors.
High temporal resolution: the event camera has microsecond-level time precision and can capture the tiniest brightness changes during high-speed motion. No matter how fast the target, the event never "blurs". A conventional frame camera captures images at 30 FPS; an event camera responds to change at millions of events per second.
Ultra-low latency: events are generated instantly, with no need to wait for a full frame to be captured. This makes event vision highly advantageous for real-time control, obstacle avoidance, gesture recognition, and similar tasks.
Very high dynamic range: because each pixel handles changes individually, event vision can operate at a dynamic range of 100 dB or higher. It stays clear even in environments where strong light and deep shadow coexist.
Data sparsity: only the changing parts trigger events, so the data volume drops sharply. A higher compression ratio means lower bandwidth and storage needs.

2. Event data vs. conventional frame images

Property	Conventional frame camera	Event camera
Acquisition mode	Synchronous frames	Asynchronous events
Temporal resolution	Millisecond (30~120 FPS)	Microsecond
Data density	Dense (all pixels)	Sparse (only changes)
Latency	High	Low
Dynamic range	Usually < 60dB	Up to 100dB+
Power	High	Lower (output on demand)
Motion blur	Noticeable	Almost none

The two are compared intuitively below — the frame camera records the whole picture (including static background), the event camera records only the moving target:

Frame image vs event image comparison

Taking a golf swing as an example, the event camera captures only the trajectory of the club and ball, not the stationary sky and grass:

Golf swing: only the motion trajectory is recorded

3. HVS: pixel-level fusion of image + event

The above describes a "pure event camera". What our camera does is HVS (Hybrid Vision Sensor) — based on the AlpsenTek HVS chip, every pixel in the same pixel array has both an image-sampling chain and an event-detection chain. The camera therefore outputs both conventional APS image frames and an EVS event stream, and the two streams are synchronized at pixel level and microsecond level. Full introduction at HVS Hybrid Vision.

Three operating modes

Mode	Output	Suited for
Image mode (APS)	Full image frames (RAW)	Static scenes, when texture/detail is needed
Event mode (EVS)	Async event stream `(x, y, p, t)`	High-speed motion, low latency, low power
Fusion mode (HVS)	Image + event dual-stream sync	When both detail and fast response are needed

Pixel-level fusion vs sensor-level fusion

There are two ways to put APS and EVS together, with very different outcomes:

Sensor-level fusion: two independent sensors (one APS + one EVS) placed physically side by side and aligned afterwards — there is parallax, the timing is not synchronized, and the package is bulky.
Pixel-level fusion (HVS / our approach): photoelectric conversion + image readout + event detection are integrated inside the same pixel, realized on-chip with architectures such as iampCOMB / GESP / IN-PULSE DiADC combined with PixMUX time-division multiplexing and BSI stacking — zero parallax, microsecond-level sync, single chip.

Dual-stream synergy: see clearly + see fast

APS and EVS each have their strengths and complement each other once fused:

APS: texture, color, static detail — "see clearly".
EVS: high-speed change, low latency, low bandwidth, >120dB high dynamic range — "see fast".

HVS vs pure EVS

Pure EVS sensors (e.g. the pure-event approach of Sony, Prophesee) pursue the ultimate event performance but usually do not output images; HVS delivers both image and event from a single chip, which is better suited to real-world scenarios that need "both a picture and events" — recognition, surveillance, interaction, and other applications that require image context.

4. How parameters affect the picture: accumulation time

The event camera itself only outputs sparse event points. To "see it as a picture", you need to accumulate the events within a time window into one frame; that window is the accumulation time:

Accumulation time too short → too few events, the picture is sparse and broken, the target is unclear;
Accumulation time appropriate → the outline of the moving target is clear;
Accumulation time too long → multiple motions overlap, the picture smears and blurs together.

How to adjust

For USB cameras, adjust the accumulation parameter in MultiVision Studio; for MIPI modules, handle frame stacking in evs_live_player on RDK X5.

The accumulation time is also coupled with the display frame rate: when the accumulation time equals the frame period (1 / frame rate), each event appears exactly once (full accumulation); when it is larger than the frame period, the same event spans multiple frames (over accumulation, appearing as slow motion / smear); when smaller than the frame period, most events are discarded (under accumulation). The trade-offs of the three schemes and XYT 3D visualization are covered in Event Visualization.

Recording ≠ visualization

The accumulation time is only a display parameter, similar to slow motion on a player. RAW / EVT2 files recorded with MultiVision Studio or evs_live_player store all the original events, independent of the accumulation time set during recording — it can be freely reset on playback. See Event Visualization.

Easy to confuse: temporal resolution ≠ latency

Temporal resolution refers to the precision of the event timestamp (microsecond-level), determined by the pixel circuit — "how finely it is recorded". Latency refers to the actual lag from when an event is generated to when it is consumed downstream, affected by transmission, processing, scheduling, etc. The two are independent: you can have microsecond temporal resolution but millisecond end-to-end latency.

5. Application scenarios

Industrial inspection: high-speed production-line defect detection
Smart transportation: fast target tracking, lane keeping
Robot navigation: stable navigation in low-light, complex dynamic scenes
Augmented/Virtual Reality (AR/VR): low-latency gesture recognition and interaction
Medical imaging: e.g. eye tracking, neural monitoring

6. Summary

Through biomimetic mechanisms and asynchronous processing, event data solves the compute bottleneck and dynamic-scene limitations of conventional vision systems. Its low power, high real-time performance, and privacy-friendly nature give it unique advantages in consumer electronics, autonomous driving, and more. As the neuromorphic computing ecosystem matures, event data will drive further advances in edge-intelligent devices.

In a conventional image-sensing system, the camera periodically captures a complete image at a fixed frame rate. Whether or not the scene changes, every frame contains the brightness values of the entire field of view. This approach is intuitive but has two fundamental limitations:

Low temporal resolution: a fixed frame rate cannot reflect rapidly changing dynamic scenes.
Redundant information: most pixels do not change significantly between adjacent frames yet are still captured repeatedly, wasting resources.

Event-based vision breaks this framework completely. It no longer "takes" pictures — it perceives change.