Detecting drones a camera can barely see
A single camera, a stream of almost-empty frames, and statistical machinery that turns a four-pixel flicker into a confident alarm, with the false alarm rate capped in advance.
The target hides below the noise
There is a real drone in this clip, about 4 pixels wide, drifting past the tower. The inset magnifies it 4x and circles it; even then it is a faint wisp. Without the circle, you would not find it.
A small drone far from the camera covers a handful of pixels, and its signal is often weaker than the sensor's own noise. Engineers call this low signal-to-noise ratio (SNR): the target is quieter than the static.
So in any single frame there is nothing to enhance, and frame-by-frame detectors, including modern neural ones, have nothing to work with.
Yet the information is there. It is just spread across time.
Don't look harder. Watch longer.
One frame says nothing, but a real drone leaves a consistent trail across many frames. The thesis turns that into a detector with two defining properties:
Patience that finds the invisible
A target too faint for any single frame still moves like a real flying object, and that consistency is information. The detector collects it, frame after frame, until a signal far below the noise becomes undeniable. Random flicker never adds up this way; a real trail does. No operator staring at a screen, and no "enhance": the accumulation is the detection.
An alarm you can hold to a number
The system watches continuously and may raise the alarm the moment it becomes sure, yet the false alarm rate stays under a budget fixed in advance. Most detectors give you a score and leave the threshold as your problem; this one ships with its error rate as part of the design, and every alarm also carries a separately calibrated "how sure are we" percentage.
"The new ground is doing both at once: patience that finds the invisible, with an error rate you can promise in advance."
Three moves, repeated every frame
Learn "normal"
The system continuously learns what this scene ordinarily does (sensor noise, clouds, buildings, slow drift) from past frames only, and keeps refreshing as the scene evolves. Whatever that model cannot explain is genuinely surprising. The fit is simple and inspectable, not a black box.
Follow the motion
Surprises only count when they line up over time the way a real flying object moves; random flickers cancel out, a consistent trail compounds. How this is done at scale, across the whole frame at once, without breaking the error promise, is the technical core of the thesis, and it stays in the thesis until it is published.
Stop when sure
The alarm fires the moment the accumulated evidence clears a bar fixed in advance by the false alarm budget. A separately calibrated head then converts the evidence into an honest "how sure are we" percentage.
Two outputs, kept apart on purpose: the alarm, whose false alarm rate is capped by construction under the modeled noise and then re-verified on real footage, and the probability, calibrated separately. And note what is absent: no neural network in this core. Every constant is inspectable, which is what makes the promise testable at all.
Measured on a frozen 200-video benchmark
Four independent public sources, two sensor families, and frozen means frozen: none of these videos is ever used for tuning or training, and every change to the system is re-measured against all of them. 195 of the 200 carry a scoreable target.
| Source | Sensor | Videos | What it contributes |
|---|---|---|---|
| Anti-UAV410 | Thermal infrared | 98 | The field's standard anti-drone benchmark; real tracking footage from pan-tilt cameras |
| Halmstad (Sweden) | Infrared | 63 | Public multi-sensor benchmark; contains both drones and birds |
| Halmstad (Sweden) | Visible light | 10 | Same scenes on an ordinary camera; the cross-sensor test |
| CST Anti-UAV | Thermal infrared | 24 | The hard set: drones of 3 to 9 pixels, exactly the thesis regime |
Beyond the benchmark, the false alarm side has been exercised on more than 8,000 clean watch windows across every campaign to date.
Four pixels. Alarm at frame 29.
Left: the real footage. Right: the machine's belief, with its running evidence score along the bottom and the alarm line it must cross.
A real 4-pixel drone drifts past an apartment tower at half a pixel per frame. The evidence crosses the alarm line at frame 29 and keeps climbing to overwhelming levels. This drone was caught in 6 of 6 windows (100%), median 29 frames, false alarm budget intact.
The same math at opposite extremes
A drone crossing at 2 pixels per frame. The wide-area engine built for fast movers alarms in 16 frames (about 0.6 s) and reads out the drone's velocity as part of the detection (cyan arrow). This motion class: 0 of 24 for the narrow-watch protocol, 20 of 24 (83%) here.
Built and tuned entirely on thermal video, then pointed at an ordinary visible-light camera with zero changes: this crosser is caught at frame 43. The math never cared about wavelength; now that is demonstrated, not assumed.
Promised vs measured
| Quantity | Promised | Measured |
|---|---|---|
| False alarms, 40-frame watches | ≤ 5% | 6 / 2,501 = 0.24% |
| False alarms, 80-frame watches | ≤ 5% | 42 / 2,501 = 1.68% |
| Detection, all evaluable videos | measured, not promised | 158 / 195 = 81.0% |
| Detection, small moving targets (thesis regime) | measured, not promised | 57 / 69 = 82.6% |
| Detection, hovering targets | known blind spot | 0 / 15 = 0% |
Every alarm ever inspected traced back to a real moving object (a bird, a pedestrian, an unlabeled second drone) or a camera event; zero confirmed pure-noise alarms at the deployed operating point. And the misses are named, not hidden: hovering and ultra-slow targets that the background model absorbs. Those blind spots drive the roadmap.
Why the next step is more real data
The regime gap
Public benchmarks are rich in big, close, trackable drones and thin exactly where this thesis lives: the hard tiny-target set is only 24 videos. Statistical confidence there is necessarily wide; more footage in the right regime tightens it.
Cameras, not physics, set the floor
Most residual alarms trace to other people's cameras: pan-tilt-zoom jumps, automatic gain steps, gaps in the labels. Measuring the true noise floor needs flights where we control the camera, the drone, and the ground truth.
A promise needs held-out proof
A false alarm cap is only convincing on data nobody tuned on. Planned: self-collected field flights, flown with two B.Sc. students, locked as verification-only. Never touched during development, opened once at the end.
Machine learning, only where it can't break the promise
Because the core is transparent, every learned upgrade can be gated: whatever inflates false alarms is rejected, no matter how good its detection looks. The first learned upgrade is already measured.
What shipped
The learned arm keeps the exact classical procedure and lets training tune its constants: how it expects targets to move, and how strongly to weigh evidence. The result stays fully readable; it learned, for example, to expect straight-line flight. Cost stated plainly: false alarms rose from 42 to 93 of 2,501 (1.68% to 3.72%), still inside the 5% budget. Two gained videos come from a sensor family the model never trained on.
What got rejected, and why that matters
An unconstrained learned noise model looked spectacular on detection and pushed false alarms to 98.7%. The pre-registered validity gate killed it in one pass. Meanwhile a learned probability calibrator, which by construction cannot touch the alarm, halved calibration error (0.161 to 0.062). Learning goes where it is safe; the gate decides.
From prototype to field-proven
Fly and verify
Self-collected field flights with two B.Sc. students: our camera, our drone, exact ground truth. Held out as the final verification set for both the detection curves and the false alarm promise.
Learn in the safe slots
Extend learning where the gate allows it: the shipped learned motion model and probability calibrator now, a validity-constrained learned background model next, aimed squarely at the hover and slow-target blind spots.
Compose the watchtower
The combined deployment (two complementary engines, one for faint slow targets, one for fast movers) is already priced on the benchmark: 1.48% false alarms, below the single-engine 1.68%, with wider coverage. Its full detection campaign is the named next measurement.