How ICCV 2025 researchers put event cameras on the drone detection problem

Magrini and colleagues at the University of Florence presented a paper at ICCV 2025’s NeVi workshop titled Drone Detection with Event Cameras that takes the question most relevant to counter-UAS sensing and asks it head-on: can an event camera detect a small drone reliably, and what does that detection loop look like in practice?

The short answer from their work is yes, and the architecture they describe is worth unpacking.

What the paper demonstrates

The researchers built a detection pipeline around a DAVIS346 event camera — a silicon retina that outputs per-pixel brightness changes asynchronously at microsecond resolution rather than full frames at fixed intervals. Their method converts incoming event streams into time-surface representations and runs a lightweight convolutional detector over those surfaces to produce bounding-box proposals.

The key result is that the event camera’s inherent motion sensitivity makes it naturally suited to the drone detection task. A small multirotor against a cluttered background produces almost no events when it is stationary. The moment it moves, it generates a distinctive burst of events from its rotating propellers and translating body — a signature that would be invisible to a frame camera against a similar background until it subtracted frames explicitly.

Detection latency in their experiments dropped well below what frame-based systems achieve, simply because the event camera does not have to wait for an integration window to close.

Why this matters for terminal-phase sensing

The drone detection use case in the paper is perimeter surveillance — detecting a drone approaching a protected site. But the sensor physics they are exploiting are the same ones that make event cameras attractive for terminal guidance.

In a terminal intercept scenario the target is closing fast, angular rate is climbing, and the guidance loop needs a centroid measurement at sub-millisecond update rates. A frame sensor integrating over 10–30 milliseconds smears the target at exactly the wrong moment. An event camera produces a measurement the moment any pixel on the target changes brightness — which, for a crossing UAS, is continuously.

The Magrini et al. paper does not address terminal geometry or the seeker integration problem, but it establishes an important baseline: detection of small rotary-wing UAS from event camera output is achievable with a lightweight inference pipeline that fits comfortably within embedded compute budgets.

The propeller signature

One observation in the paper worth highlighting: the propeller tips, which move significantly faster than the drone body, produce a characteristic event density pattern distinct from background clutter. Propellers on typical Group 1–2 UAS spin at 4,000–8,000 RPM, which means tip events repeat at rates that frame cameras cannot resolve but event cameras see clearly.

This is a signature that frame-based detectors cannot exploit without extremely high frame-rate sensors that come with serious bandwidth penalties. It is available to event-camera pipelines for free.

Open questions

The paper evaluates detection under relatively benign outdoor conditions. The harder problems — high-altitude targets against bright sky, targets in rain or haze, multi-target disambiguation — are left as future work. Those are the conditions that matter most for the C-UAS mission, and they are the direction we expect the literature to move over the next year.

Tracking the propeller signature through occlusion and variable lighting will be one of the more technically interesting challenges in that extension.

---

Magrini et al., “Drone Detection with Event Cameras,” ICCV 2025 NeVi Workshop. arxiv 2508.04564