Optical Counter-UAS: The Harsh Physics of Drone Detection at 2 Kilometres

A DJI-class quadcopter measures roughly 35 centimeters across. Place it two kilometers away, track it with a 4K camera using a 110 mm lens (roughly a 6° horizontal field of view on a 1/1.8″ format), and that drone registers on your sensor as roughly six pixels.

Six pixels. That is the entire evidentiary basis a counter-UAS (C-UAS) system relies on to decide whether to wake a battery commander at 03:00.

At Obraz, we engineer edge-deployed optical computer vision systems for contested and mission-critical environments. This article examines:

• The economics of passive sensing
• The optical physics that govern detection
• The mathematics of false alarms
• Practical edge deployment
• An evaluation framework for assessing optical counter-UAS systems

Most literature on AI in air defense skips past this mathematical reality and jumps straight to the economics. Those economic arguments are correct, but they conflate two different threat classes. A $500 FPV quadcopter and a $20,000+ long-range one-way-attack (OWA) drone like the Shahed-136 pose very different sensing problems at very different ranges. The economics are the easy part either way. The hard part is that the underlying sensing problem is a genuinely difficult exercise in optical physics, signal statistics, and false-alarm arithmetic. It does not yield to a fine-tuned YOLO checkpoint and a flashy press release.

Here is a technical account of what actually breaks, what actually works, and, at the end, an engineering checklist you can use to rigorously evaluate any vendor’s claims, including ours.

1. Why the Economics Force a Passive Answer

The cost-exchange argument has become a cliché, but it is entirely true [1].

• Russian-operated Shahed-136 / Geran-2 one-way attack drones cost in the tens of thousands of dollars.
• A PAC-3 MSE interceptor costs on the order of four million dollars [2].
• In the Red Sea, Western navies have expended SM-2 and SM-6 interceptors (worth $2M – $4M+ each) against drones worth a fraction of a rounding error [3].
• In Ukraine, first-person-view (FPV) drones costing a few hundred dollars routinely destroy armor worth millions, with both sides consuming tens of thousands of drones monthly [4].

The usual conclusion is “we need cheaper effectors,” which is only half right. Cheap effectors already exist: guns, electronic attack, high-power microwave, interceptor drones, and directed energy. Legacy 40 mm Bofors L/70 and 23 mm (ZU-23-2) gun systems have been heavily utilized by Ukrainian mobile fire teams against Shahed drones precisely because their cost per engagement is trivial, preserving expensive interceptor missiles for high-value threats.

The problem with cheap effectors is that they are short-ranged, “dumb,” and require an accurate, verified, continuously updated track to be worth firing. A gun with a bad track is worse than no gun: it produces wasted expenditure, collateral risk, and zero kills.

A modern C-UAS sensor must:

• Be passive: it does not radiate, so it cannot be geolocated by RF direction-finding or targeted by anti-radiation weapons.
• Be independent of emissions: it cannot depend on the target emitting signals.
• Be economically scalable: it must be produced at a unit cost consistent with covering hundreds of dispersed sites.
• Classify, not just detect: knowing “something is out there” does not authorize a kinetic response.

This points directly to electro-optical/infrared (EO/IR) sensors backed by competent computer vision.

2. What Recent Conflicts Actually Proved

• Ukraine demonstrated volume. The airspace is permanently saturated with small airframes. Any drone detection system with a false-alarm rate that is merely “tolerable” in a lab will generate severe operator fatigue.

• Nagorno-Karabakh (2020) demonstrated sensor deception. Azerbaijani forces repurposed unmanned An-2 biplanes as decoys to bait radar emissions and reveal Armenian firing positions [5]. Radar-only air pictures are deceivable.

• Ukraine and the Black sea (2024-2026) demonstrated synthetic radar deception.  Russian forces routinely deploy Parodiya and Gerbera foam decoys equipped with 3D-printed Luneburg lenses. These $400 gliders artificially amplify their Radar Cross-Section (RCS) to mimic armed Shahed-136 attack drones on radar, deliberately baiting multi-million-dollar interceptors. Radar alone cannot tell foam from a 50 kg warhead—only high-resolution optical classification can.

~ Ukraine and the Black Sea (2024–2026) demonstrated synthetic radar deception.

• The Middle East and Red Sea demonstrated the denied environment. One-way attack drones rely primarily on inertial navigation plus jam-resistant GNSS.

~ Indian Army’s Upgraded L-70 Guns Decisively Repel Pakistani Drone Swarm over Jammu,May 2025

• The India-Pakistan exchange, May 2025, demonstrated density and ambiguity. Cross-border incursions over populated areas generated significant ambiguity, and many specific claims from both sides remain contested. What is not contested is the operational requirement: distinguishing a hostile UAV from a civil aircraft, a bird, or a festival balloon, autonomously, at night, in seconds, at scale, with an engagement decision that survives rigorous scrutiny.

3. The Physics You Cannot Argue With Detection

What is Instantaneous Field of View (IFOV), and why does it decide detection range?

Everything in long-range optical detection stems from one foundational metric: the angular size of a single pixel.

IFOV = HFOV / horizontal_pixels

Take a 4K sensor (3,840 px) behind a 110 mm lens (roughly a 6° horizontal field of view on a 1/1.8″ format). This yields an IFOV of 27.3 μrad per pixel. A target’s apparent size is its critical dimension divided by range, divided by IFOV.

Table 1: Optical C-UAS Pixel Density by Target Size and Range (4K sensor, 110mm lens)

This table is the ultimate lie detector. If a vendor claims they can identify an object at 5 kilometers, ask for their IFOV.

Detection ≠ Recognition ≠ Identification

Under one commonly cited formulation of the Johnson criteria, discrimination thresholds sit at roughly 1.0 line pair for detection, 4.0 for recognition, and 6.4 for identification [6]. (Different sources use different exact multipliers; the ordering, not the precise constants, is what survives across formulations.) Even so, the ranges differ by vast factors, not percentages. A system that “detects” a target at 3 km may not recognize whether it’s a quadcopter or a bird until roughly 1.6 km, and may not identify it until 1.0 km.

How the Atmosphere Sets a Hard Ceiling

Optical range is strictly bounded by the medium:

• Turbulence: convective conditions over hot ground smear targets.
• Extinction: rain, fog, and dust attenuate contrast exponentially.
• Thermal crossover: around dawn and dusk, target and background temperatures converge, collapsing infrared contrast.
• Solar geometry: targets on a low-sun bearing sit in a glare field.

4. Why False Alarms Are the Real Bottleneck

To find sub-ten-pixel targets in a 4K frame, you must slice it into overlapping tiles and run inference at native resolution. Slicing a 4K frame yields roughly 40 tiles. At 30 fps, that requires 4.32 million tile-inferences per hour, per camera.

If a detector produces a spurious detection on just one tile in ten thousand, the result is 432 false alarms per hour. On a twelve-camera site, that is close to one false alarm every second. To hold false alarms under one per hour on a twelve-camera site, the per-tile false-positive rate has to drop below roughly 1.9 × 10⁻⁸.

Detect over time, not over pixels

Requiring a candidate to persist across multiple frames within a kinematically plausible gate makes uncorrelated sensor noise fall away. But this only works against uncorrelated false alarms. A soaring raptor has a persistent, kinematically plausible track. Voting doesn’t remove these. Only classification does, and the classifier only works if those exact categories are in the training distribution.

These physical constraints are universal. The question is not whether they can be ignored, but how an engineering team designs around them in practice.

5. How Obraz Engineers Optical C-UAS Systems

• In-House Data Security:  – We run a 200+ person internal annotation workforce. Because we do not outsource to third-party clouds, your defense data remains strictly on-premise, secure, and confidential.
• Multi-frame detection: Candidate evidence accumulates across a frame buffer before detection is declared.
• Resolution preservation: We use high-resolution-preserving backbones (FPN and HRNet) combined with native-resolution tiled inference.
• Kinematic discrimination: Soaring raptors don’t flap, making them the hardest negative. We extract both photometric oscillations and trajectory-shape statistics.
• Point annotation: Human annotators can’t draw an accurate bounding box around a five-pixel object. We annotate extreme-range targets as center points and derive extent analytically, removing annotator drift.
• The data problem: Public drone datasets are inadequate for this. We rely on purpose-built field collection, physics-based synthetic augmentation, and systematic hard-negative mining from deployed systems.
• Honest monocular range: We solve range ambiguity with stereo baselines and sensor fusion, never presenting monocular range as an absolute measurement.

6. Edge Inference With Real Constraints

Backhauling uncompressed 4K video breaks tactical bandwidth. Inference has to happen at the edge. A modern edge AI module like the NVIDIA Jetson AGX Orin, or the newer Jetson Thor (T2000 series) which became available in 2026, delivers INT8/FP4 throughput in the hundreds of TOPS within a configurable 15-60 W envelope [7]. In the field, the constraints are thermal and power. Sustained throughput under thermal throttling at 48°C ambient inside a sealed IP67 enclosure is the only number that matters, not the datasheet peak.

7. Sensor Fusion for Optical Counter-UAS Verification

Radar is not obsolete. The correct anti-drone architecture is complementary:

1. Radar cue: wide-area detection.
2. RF detection: free identification where a control link exists.
3. Acoustic sensing: non-line-of-sight detection of rotor harmonics in urban canyons.
4. EO/IR slew-to-cue: narrow-field, precision angular tracking and strong classification.
5. Human decision: the fused, classified track goes to an operator who retains engagement authority.

By integrating our inference layer with existing perimeter CCTV and legacy EO masts, installed capital becomes usable small-UAS-detection capability at a fraction of the cost of a greenfield deployment.

Conclusion

The sky over a defended site isn’t won by slogans about asymmetry. It’s won by getting the optics right, being honest about atmospheric limits, engineering for false-alarm rates, and fusing modalities to prevent deception.

Six pixels is not much to work with. But it is enough for a well-engineered optical counter-UAS system, provided you treat it as a physics and statistics problem rather than a machine-learning demo. At Obraz, we’re building the visual cortex of modern air defense: passive, edge-deployed, honestly characterized, and designed to make cheap effectors accurate enough to be worth firing.

The Evaluation Checklist

If you take one thing from this piece, use this checklist to rigorously evaluate any optical counter-UAS vendor, including us.

• Ask for IFOV, not detection range. Demand the angular resolution per pixel.
• Demand the false-alarm rate in false tracks per hour, per camera, measured over at least 100 hours on an operational background.
• Require the full DET/ROC curve, not a single cherry-picked operating point.
• Specify harsh conditions in the test plan: dawn/dusk thermal crossover, low-sun glare, rain, and haze.
• Put live birds in the frame. Soaring raptors are the hardest negative; a vendor who hasn’t tested against them hasn’t tested their system.
• Measure photon-to-track-message latency end to end, including slew-to-cue settling time.
• Test the AI system on the actual hardware that will be deployed in the field, after it has been running inside its real protective enclosure under realistic environmental heat conditions for four hours. Do not measure performance only on a powerful computer/GPU in a laboratory.
• Ask how range is estimated. If it’s monocular and lacks a stated uncertainty, it’s unreliable.
• Ask for failure modes. A team that can’t list five conditions where its system degrades has never deployed it.

We’ve distilled these engineering principles into a practical evaluation matrix that procurement teams, systems integrators, and defense organizations can use to assess any optical counter-UAS solution, including ours. 

References & Citations

[1] Batmanghelidj, Esfandyar. “Drones Like Bicycles: The cost of a Shahed-136 kamikaze drone.” Phenomenal World, March 11, 2026.

[2] Center for Strategic and International Studies (CSIS). “Patriot Advanced Capability-3 (PAC-3).” Missile Defense Project, 2024.

[3] LaGrone, Sam. “U.S. Navy Has Fired $1 Billion in Munitions against Houthi Rebels.” USNI News, April 2024.

[4] Panella, Chris. “How $400 FPV Drones Are Changing the Economics of War.” Business Insider, 2024.

[5] Shaikh, Shaan, and Wes Rumbaugh. “The Air and Missile War in Nagorno-Karabakh: Lessons for the Future of Strike and Defense.” CSIS, Dec 8, 2020.

[6] Johnson, John. “Analysis of Image Forming Systems.” Proceedings of the Image Intensifier Symposium, US Army Engineer Research and Development Laboratories, 1958.

[7] NVIDIA Corporation. “NVIDIA Jetson AGX Orin Technical Specifications” (2022) and “Jetson Thor T2000/T3000 Series Launch Specifications” (2026).