Anti-UAV Detection

A real-time anti-UAV (drone) surveillance system and the research paper that evaluates it honestly. The headline isn't "we built a detector" — it's "we measured what actually survives when the data changes." Co-authored with Rawad Hammad at the University of East London, written up as an IEEE-format manuscript.

A three-stage cascade: a YOLO11s detector finds candidates, a ResNet-18 verifier rejects birds, and an EfficientNet-B0 classifier names the drone type. On top sits a Streamlit surveillance app — image and RTSP/IP-camera modes, restricted-zone alerts, multi-drone tracking, and an alert database.

In-distribution, every stage looks almost solved: 0.97 mAP@0.5 detection, ~100% bird suppression, 99.7% balanced classification accuracy. That's exactly the kind of near-saturated number that should make you suspicious. The paper asks the only question that matters for a real deployment: do those scores hold on data the models have never seen?

They don't hold — but they fail in different ways, and that structure is the contribution:

• —The classifier generalises — 92–99% on out-of-distribution drone types.
• —The detector loses about half its mAP — 0.97 → 0.52 cross-domain, replicated on a second benchmark (0.55), and the drop tracks target size.
• —The verifier collapses — from ~100% bird suppression to ~26% on novel bird species it was never trained on.

So the weakest link isn't where the in-distribution scores suggest. A system that looks uniformly excellent is, under domain shift, only as trustworthy as its most brittle stage.

Before trusting any of it, I ran a perceptual-hash (pHash) leakage audit to catch near-duplicate images bleeding across the train/test split — and it materially changed the numbers. Everything downstream is reported behind that audit, multi-seeded (3 seeds), with frozen split manifests and a fixed zero-shot evaluation convention so the cross-domain results are reproducible rather than cherry-picked.

Different domain from my finance work — computer vision, not markets — but the same instinct runs through it: distrust a number that looks too good, find where the system actually breaks, and report it plainly. The paper source and evaluation code are on GitHub.