事件相機走出實驗室：神經形態視覺在倉庫、汽車與工廠的應用

The most annoying number in robotics is thirty-three milliseconds. That is the delay a 30-fps camera imposes before any computation begins, purely because it has to wait for the next frame to be captured and read out.

For a robot arm picking a moving object off a conveyor, 33 ms of blind time is the difference between a clean pick and a jam. For a car detecting a pedestrian stepping out, it is a metre of travel at motorway speed. For a warehouse AMR turning a corner, it is the reason you have to slow down at intersections.

Event-based vision — the sensor family that grew out of neuromorphic engineering — attacks that number directly. Instead of capturing full frames on a clock, each pixel independently reports a change in brightness the moment it happens. No frame clock, no quantisation delay, and no data for the parts of the scene that are not changing.

The latency argument, quantified

The comparison that matters is not sensor latency in isolation. It is the whole loop: capture, transfer, preprocess, infer, actuate.

A frame camera at 30 fps contributes 33 ms of quantisation delay on average, plus readout and transfer. Add a GPU-based detector and you are at 45-50 ms before the actuator even hears about it. An event sensor with an on-sensor neural network can get the whole loop under 2 ms, because there is no frame boundary to wait for and the sparsity means the compute is proportional to what changed, not to how many pixels exist.

Eighteen times faster is the kind of number that changes a design rather than optimising one. It is the difference between "slow down at intersections" and "do not slow down".

Why it stayed niche for fifteen years

Three reasons, all of them engineering, none of them fundamental:

No frames means no CNNs. The entire deep-learning toolchain is built on dense tensors. Event data is sparse, asynchronous, and has a time dimension. Porting a ResNet does not work; you need spiking networks, or event-frame accumulation (which throws away the latency advantage), or graph representations.
Data was scarce. Training needs datasets, datasets need sensors, sensors need a market. The loop took a decade to close.
Tooling was hostile. No OpenCV, no standard format, no simulator. The first cohort of researchers wrote their own everything.

All three are now unblocking. Event datasets exist at useful scale, spiking and event-native network architectures have matured, and the sensor vendors ship SDKs that do not require a PhD.

The market is small and compounding

Event-based vision is roughly 1.7% of industrial vision spend — a rounding error next to the frame-based market, which is mature and enormous. But the growth rates are completely different: mid-teens for frame-based, low-twenties for event-based.

Three demand pockets are doing the pulling:

Automotive driver monitoring. Detecting a driver's eye movement and blink pattern at 200 Hz, in a dark cabin, at a power budget a car can afford. Frame cameras struggle with the dark; event sensors do not care about absolute brightness, only change.

Warehouse robotics. Fast-moving objects, tight tolerances, and a hard requirement for deterministic latency. Event sensors give the control loop a fighting chance at sub-10 ms response without a GPU.

Industrial inspection. Detecting a defect on a part moving at 2 m/s. A frame camera blurs it; an event sensor sees the edge as it passes. This is the use case with the clearest ROI and the least marketing noise.

What changes in ordinary life

Machines stop hesitating. The uncanny pause before a robot reacts — the half-second where a cobot arm stops because a person walked into its cell — is largely a sensing latency artefact. Removing it makes automation feel less like automation and more like a colleague with decent reflexes. That sounds cosmetic; it is actually the main barrier to deploying robots near people.

Cars get better at the edge cases that kill people. Pedestrian detection at night, cyclist detection in rain, driver drowsiness on a motorway. Event sensors are not a replacement for cameras and lidar — they are a fourth modality that happens to be very good at exactly the fast, dark, high-contrast situations where the others are weakest.

Power budgets drop in always-on sensing. An event sensor in a battery-powered doorbell or security camera draws microwatts when nothing is happening, because nothing is happening is the point. That turns always-on vision from a battery problem into a non-problem.

A privacy side effect nobody planned. Event data is sparse and intensity-based, not photographic. You often cannot reconstruct a recognisable image from it. That makes event sensing a genuinely better fit for always-on monitoring in homes and workplaces — not because it is more private by design, but because it throws away the information that makes images identifying.

What to watch

On-sensor inference, not sensor-to-GPU pipelines. The latency win only materialises when the network runs on the sensor stack. Vendors who ship "event sensor plus host GPU" are selling a faster camera, not a faster loop.
Event-native model architectures with real benchmarks. Watch for published latency/accuracy curves on standard tasks, not demos.
Cost per pixel. Event sensors are still expensive per pixel compared to CMOS. The crossover to commodity pricing is what turns this from a niche into a default.

The engineering summary

Event-based vision is not a better camera. It is a different contract with the world: report change, not appearance. That contract is worse for photography and dramatically better for control.

The industry spent fifteen years waiting for the software stack to catch up with the sensor physics. It has, roughly, and the consequence is that the fastest-growing segment of machine vision is now the one that does not produce images at all.

If this was useful

Digging through the benchmarks and building these comparisons takes real effort. If it saved you time, a tip is genuinely appreciated.

SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS

Creating this takes real effort — thank you for the support.