Scotopic Request a briefing

Hold the track. When the link is gone.

Detection at the sensor · 26 bytes per track update · No guidance, no weapons

Scroll

Jam the link, and most ISR payloads become dead weight. The intelligence was never at the sensor.

One hard problem, done properly.

Detect

Small aerial targets at range, against sky, terrain and clutter. A drone at 2 km is a handful of pixels: at that scale shape carries almost no information, and most detectors quietly stop working. This is the part we built for.

Track

Kalman tracking with coasting through occlusion, so a target that passes behind terrain is picked up again rather than lost. Motion is what separates a drone from a bird when shape cannot.

Report

A versioned track contract with bearing, elevation, size and per-component uncertainty. Buffers on link loss, flushes losslessly on reconnect. The human makes the call.

Detect Track Report Human decision

System architecture

EO/IR SENSOR ONBOARD PERCEPTION DETECT → TRACK REPORT TRACK + STATE OPERATOR HUMAN DECISION

Perception and tracking run at the sensor. Guidance-free. The operator makes the call.

MODULE · FOVEA

FOVEA

The perception module. Detect, track, report. At the sensor, without GPS, without a link.

TRL 4 overall · self-assessed. Edge deployment, output contract and packaging at TRL 5; detector and tracker at 4. SAPIENT interface in progress.

Measured

Target hardwareRaspberry Pi 5 (8 GB) + Hailo‑8 · 26 TOPS
Whole‑frame rate173.6 ms per 4K frame, 5.76 FPS · full field of view, continuously
Inference latency, CPU595 ms p95 · no accelerator, 1280×1280, sustained, 28 Jul 2026
ThermalChip 49.5 °C, host 50.2 °C after 10 min sustained · sustained run, no measurable drift · bench, open air, mains power
PortabilityFull suite green on aarch64 · 0 failures
Recall · small targets86.5% · held‑out split, in‑distribution. Size band and match criterion in the method note.
Recall · cross‑dataset57.2% · zero‑shot, unseen dataset. The honest number, published because it is the one that matters.
False alarms · own footage0.098 per frame · target‑free sky, zero‑shot, measured through the shipped build. Denominator and conditions in the method note.
Size floor8 px minimum resolved size · below it the detector enters a statistical regime and most targets are missed. Range follows from your optics and sensor, not from our software — we publish the threshold so you can compute yours. Most vendors publish a range and not a threshold.
ThresholdCalibratable · P/R/FAR sweep, knee published
Track state on the wire26 bytes per update · compact binary, lossy by design, every field pinned to documented precision by documented precision test
Sustained rate2.1 kbps at 10 Hz, one track · 7.8 kbps at five tracks
UncertaintyBearing and elevation carry their own 1σ · drops into a fusion filter without being re‑derived
Link resilienceBuffered on loss · lossless flush on reconnect
RuntimeONNX · at the sensor, no cloud, no licence server

In progress

Track robustnessReacquires after occlusion in synthetic scenario · not yet flight‑validated
Field validationSelf‑collected flight footage · Aug 2026
Detection rangeNot characterised in metres · pixel floor above, but no surveyed‑truth range trial yet
Bearing accuracyNot yet characterised against surveyed truth
Mass and powerNot yet measured in enclosure, on battery
SAPIENT interfaceBSI Flex 335 · implementation planned, not validated against a certified DSM

MEASURED = reproduced on the hardware named above, or on a held-out split where stated, on a stated date. IN PROGRESS = not yet measured, published when it is, whatever it shows. Software module today: no sealed hardware unit.

Measured performance

26 B

per track update, on the wire
compact binary, size pinned by test
2.1 kbps sustained at 10 Hz

5.76 FPS

the whole 4K field of view, every frame, at the sensor
nothing has to emit, and nothing has to point us at it
10 min sustained · no throttling

Every figure on this page was reproduced on the hardware it names, on a stated date. The numbers we have not taken yet are listed as such, with the same weight.

Method and raw logs available on request

Video downlink13.16 Mbps · HEVC 992×1736, ffprobe
HEVC 992×1736 · measured with ffprobe on real footage
Track state, JSON64.7 kbps
Track state, compact2.1 kbps · 26 B at 10 Hz
6,327 : 1 Bars are to scale. The compact bar is drawn at its minimum visible width and is still an overstatement.

Accelerated figures measured 13 Aug 2026 on the target hardware named above; the 595 ms CPU baseline measured 28 Jul 2026 on the same board without the accelerator. False-alarm figures measured 13 Aug 2026 through the shipped build on target-free sky, every denominator recorded. Recall of 86.5% is a held-out split, in-distribution; the cross-dataset figure of 57.2% is on the same page for the same reason. Field validation on self-collected footage is under way and will be published measured, whatever it shows.
All footage on this page: system demonstration, synthetic scenario.

A component, not a replacement.

There is real work happening in autonomy right now, and some of it is very good. Almost none of it is aimed at the specific problem of finding a small aerial target at range, on cheap compute, with a false-alarm rate an operator will tolerate. That is the whole of what we do.

Autonomy stacksReplace the flight software. You adopt their operating system, their board, their qualification path, and you become a reseller of someone else’s aircraft.
ScotopicSits next to what you already have. Your autopilot, your radio, your ground station. It reads a camera and writes a track. Nothing else changes.
Full counter-UAS systemsRadar, effector, command and control, sold as one procurement to a ministry, priced accordingly. Excellent, and out of reach for a platform maker who needs one capability.
ScotopicThe optical detection layer on its own, on roughly €200 of commercial compute. Onboard an aircraft, or on a mast looking up.
Video downlink and ground analyticsThe intelligence lives on the ground. It works beautifully until someone jams the link, which is the moment you needed it.
ScotopicThe intelligence is at the sensor. What goes down is 26 bytes, buffered through the blackout and flushed losslessly on reconnect.
Building it yourselfAn architecture anyone can download. Then six to twelve months on quantisation, thermal limits, false alarms and, above all, a dataset nobody hands you.
ScotopicAlready built, already measured on the target hardware, with the failure envelope written down rather than smoothed over.

A number is worth what the source is willing to publish beside it. That is why the cross-dataset figure and the uncharacterised rows sit on this page next to the good ones, in the same type. If you would rather not take ours on faith, run the evaluation on your own footage and read what it says. No NDA, no call.

Drops in. Changes nothing else.

Scotopic FOVEA takes your camera feed and returns a track. It does not touch your autopilot, your navigation or your radio, and it keeps seeing when the signals are gone. Onboard an aircraft, or fixed on a mast.

Form

A software perception module today. A sealed sensor and compute unit is on the roadmap, not built. We will say so until it exists.

Compute

Commercial parts, measured today. Folding the accelerated path into the shipped build is in progress; the container runs CPU-only until it lands. no custom silicon, no cloud dependency and no licence server. Roughly €200 of compute per node. It scales up if you have more; it was built where there is less.

Navigation

Nothing external in the loop. Detection and tracking depend on the sensor and the compute next to it, not on GPS, the datalink or a ground station. Track output is relative: bearing, elevation, size, each with its own uncertainty. World-frame geolocation needs host position and attitude from your platform, and we report the error that adds.

Interface

One versioned track contract, regardless of platform. Compact binary on the wire at 26 bytes today, structured JSON for integration and debugging. Bearing and elevation carry decomposed uncertainty, so the output drops straight into your filter instead of being re-derived. MAVLink and ROS 2 bindings on the roadmap. Full spec on request.

Why software, and why us.

Scotopic started in robotics. Building machines that move is a good problem, but after a while the interesting part stops being the mechanism and starts being the judgement: what the machine sees, what it makes of it, and what it hands back to the person responsible.

That question got sharper when it stopped being academic. On this flank of Europe the airspace is contested now, and the aircraft flying in it go blind the moment someone decides to jam them.

Hardware is where you get locked in. Software is where you stay free. A perception layer is not tied to one airframe, one vendor or one supply chain. It runs on parts anyone can buy, on aircraft that already exist, and it makes them better without replacing them. That is the whole thesis: do not build another platform, build the thing that gives the platforms already in the air something worth having.

As for why us. We are small, and small is the wrong reason to buy from someone, so here is the right one. We chose one narrow problem, small aerial targets at range on cheap compute, and we measure it the way it deserves. When our validation set turned out to be flattering us, we rebuilt it and wrote down that every number before that line was wrong. That habit is the product as much as the code is.

Scotopic is built in Romania, on the eastern flank, against the threat picture that is actually here. It is deliberately ISR only: detect, track, report. No guidance, no weapons, no autonomous engagement. The decision stays with a human, by design and not by accident.

EntityAERO-METRIC INTELLIGENCE
2025 S.R.L.
BaseBucharest, Romania · EU / NATO
Founded2025
StageTRL 4 · self-assessed
Supply chainCOTS · ITAR-free
ScopeISR only · no guidance, no weapons

09 · Contact

Request a briefing.

Request a briefing

traian@scotopic.eu

For integrators, programmes and investors. Ask for the method note, or for the evaluation kit to run on your own footage.

Applicant · NATO DIANA 2026