Leonardo's event camera drone detection experiments — an industry perspective on what works in the field

Academic event-camera research typically reports on carefully controlled indoor sequences or outdoor scenarios staged to have clean backgrounds and well-characterized target motion. Leonardo’s paper (arxiv 2409.16099) is different: it’s a defense integrator’s report on what happens when you take commercial event cameras outside, point them at real drones, and evaluate them against production C-UAS requirements.

That perspective is valuable precisely because it doesn’t cherry-pick conditions. The paper reports what works, what doesn’t, and why — which is more useful for seeker engineering than a result that optimizes one metric in a controlled setting.

The experimental setup

Leonardo tested commercial event cameras — Davis346 and Metavision EVK4 — against a variety of small UAS platforms under real outdoor conditions. Targets included a DJI Phantom 4, a fixed-wing racing drone, and a custom multirotor. Ranges varied from 50 to 400 meters. Background conditions included clear sky, partly cloudy sky, tree lines, and building facades.

The paper evaluates detection performance as a function of range, target angular size, background complexity, and illumination. It’s a systematic characterization rather than a pitch for a specific algorithm — which makes it more useful as a system design reference.

What works

Against a clear sky background at ranges below 200 meters, commercial event cameras achieve detection rates above 90% for all target types with simple threshold-based algorithms. The event signature from a multirotor drone — rotating propellers that modulate the target’s apparent contrast — produces a distinctive event pattern that simple algorithms can find reliably in a clean background.

At longer ranges (300–400 meters), detection rates drop. The target subtends fewer pixels, the event rate from the target decreases, and the signal-to-noise ratio against background clutter falls. The authors report that a GNN-based detector maintains useful detection rates at ranges where the threshold detector has failed, which motivates the more sophisticated algorithm work that academic papers pursue.

What doesn’t work

Against complex backgrounds — a target crossing a tree line, or at low elevation with buildings in the background — the false alarm rate rises substantially with simple detection approaches. The tree line produces events from wind-induced leaf motion that are difficult to distinguish from target events at range. Building facades in sunlight produce specular glint events that the simple detector can’t filter.

The fix in both cases is trajectory consistency: real targets move on smooth trajectories; clutter events don’t. A tracker that requires a new detection to be consistent with the track hypothesis rejects most false alarms from complex backgrounds. But a tracker that requires multiple detections before confirming a track has higher latency than a single-frame detector, which is a cost in the terminal guidance scenario.

The bias tuning result

One practically important finding: the optimal bias settings for the event camera vary significantly with background complexity. The bias parameters — particularly the contrast threshold that determines how much brightness change generates an event — need to be higher in complex backgrounds (to suppress clutter events) and lower in clean backgrounds (to maintain sensitivity to small targets). The authors implement an automatic bias adjustment loop that monitors the background event rate and adjusts the contrast threshold to maintain a target event rate. This is the kind of operational adaptation that makes the difference between a lab result and a field-deployable system.

The system-level lesson

Leonardo’s paper is essentially confirming that event cameras work for drone detection within certain conditions and need adaptation outside those conditions. The conditions where they work well — clear sky, ranges below 200 meters, crossing targets — correspond directly to the terminal phase of a C-UAS engagement. The conditions where they struggle — complex backgrounds, long range — correspond to the search and cue phase, which is typically handed off to radar or other sensors before the terminal seeker takes over.

For a seeker design that activates in the final 200 meters against a target already cued by a radar or EO tracker, Leonardo’s results suggest that event-camera detection is reliable enough to close the guidance loop. The remaining challenge is making the bias adaptation and tracker initialization fast enough to handle the short engagement timeline.