Commercial off-the-shelf event cameras for space surveillance — the Optica result that started the conversation

Two years ago we started tracking the event-camera literature in earnest. This post closes the loop on that reading — we’re covering the paper that, more than any other, convinced us that event-based sensing had crossed a threshold from interesting research to serious operational tool.

The Optica paper on commercial off-the-shelf event cameras for space surveillance reports results from a ground-based observation campaign using unmodified Sony IMX636 and Prophesee Metavision sensors to detect and track LEO satellites. Not laboratory demonstrations with careful staging. Actual sky observations, with real atmospheric turbulence, pointing errors, and background noise from city lights.

What the paper demonstrates

The detection problem: a CubeSat at 500 km altitude has an apparent angular velocity of roughly 1 degree per second relative to a ground observer. A frame camera at video rate (30 fps, 33 ms exposure) would need to image for 33 ms while the target moves half an arc-minute. If the pixel scale is 1 arc-second per pixel, that’s 30 pixels of motion blur — the target is unrecognizable as a point source.

An event camera at the same pixel scale sees the target as a moving point that triggers events as it crosses each pixel. No blur. No accumulation of shot noise from a long exposure. The events trace out the track directly. The Optica paper detects satellites this way down to an apparent magnitude of +11 — significantly fainter than what a frame camera of comparable aperture can detect in a single short exposure.

The detection latency result: the paper reports detection latency of below 1 ms from the satellite entering the field of view. For space surveillance this is interesting mostly as a system property; the positional accuracy matters more than the detection latency for orbit determination. But the latency demonstration is what made us think carefully about terminal guidance applications, where 1 ms latency has direct guidance significance.

The COTS aspect

The “COTS” in the paper title matters. The sensors used — IMX636 and EVK4 — are commercially available sensors that any qualified buyer can procure. The paper uses them without modification, with vendor-provided drivers and bias settings that were then optimized for the space surveillance application. This is a proof that event cameras are ready for operational use, not just controlled experiments. If unmodified COTS hardware achieves these results on a challenging real-world detection problem, the hardware is mature enough to build on.

The contrast with where event cameras were five years earlier is significant. The first-generation DAVIS240 sensors required careful bias tuning and custom readout electronics and produced usable results only in favorable conditions. The fourth-generation IMX636 works reliably out of the box and has been deployed by multiple defense contractors and academic groups in the past year without requiring sensor-level customization. That maturation trajectory continues — the next generation sensors are already sampling.

Looking back over two years

Reading two years of event-camera literature continuously, from this paper forward, reveals a clear trajectory. The hardware has matured: pixel pitch, dynamic range, and noise floor have all improved, and manufacturing yields have improved enough that sensors are actually available when you want to buy them. The algorithms have advanced faster: the move from hand-crafted event filters to GNNs to asynchronous processing to SSMs represents three generations of architectural evolution in two years. The datasets have expanded: there are now labeled datasets for drone detection, automotive scenes, space surveillance, and gesture recognition that enable systematic algorithm benchmarking.

What hasn’t kept pace is the integration work — taking the algorithms from the papers and putting them into flight-qualified hardware with calibrated sensors, characterized software, and validated performance against defense-relevant targets. That gap is where we’ve been spending our engineering effort. The papers map the capability frontier; the integration work is what moves the frontier from research to product.

The next two years will be more interesting than the last two. The sensor technology is far enough ahead of the integration work that significant advances in field-deployable capability are achievable without waiting for new sensor generations. The architecture questions have been answered at the research level. The remaining questions are engineering questions, and those have answers.