Blink and It's Over: The Fatal Flaw of the Scanning Eye

Count the drones in the large attacks of the last three years. On 9 January 2024, US Central Command reported that a single Houthi attack in the Red Sea involved 18 one-way attack drones, two anti-ship cruise missiles and one anti-ship ballistic missile (Al Jazeera, citing CENTCOM). Iran's attack on Israel in April 2024 reportedly included around 170 drones along with cruise and ballistic missiles (CSIS). On the night of 8–9 July 2025, Ukraine's air force reported 728 Shahed-type and decoy drones in one night (Military.com).

Most of those drones were handled by long-range air defence, not by the short-range electro-optical sensors this post is about. But the shape of the threat has changed. It now arrives as many objects, from several bearings, at the same time. That raises a design question for anyone building the optical layer of a counter-UAS system. Does the sensor look at the whole sky all the time, or does it look at one part of the sky, then the next?

This post argues that the second approach, which we will call sequential looking, is a poor fit for the primary detection layer against swarms. Two familiar sensor classes work this way: the pan-tilt-zoom (PTZ) camera, and the scanning panoramic thermal imager. Both are good products for the jobs they were designed for, and the scanners in particular are serious counter-UAS sensors. The argument here is about physics and architecture, not about whether any one product is well made.

Two ways to look at the sky

A staring sensor points fixed optics at a fixed piece of sky and reads every pixel in every frame. To cover more sky you add more sensors. A sequential sensor moves its field of view. A PTZ camera slews and zooms to a point of interest. A scanning panoramic imager rotates its head, or a mirror, and builds a 360° image from strips over each rotation.

The difference that matters is the revisit interval: how long a given direction goes unwatched between two looks. For a staring sensor it is one frame period. For a sequential sensor it is however long the sensor takes to come back.

PTZ: a follow-up sensor, not a search sensor

PTZ datasheets are clear about the trade. A heavy-duty Axis PTZ, the Q6225-LE, gives a horizontal field of view of 63.8° at the wide end and 2.2° at the narrow end, with pan speeds up to 150°/s (Axis datasheet). FLIR's Ranger HDC MR, a defence surveillance system, has a cooled thermal channel with a 3.0° × 1.7° narrow field of view on one model, on a pan-tilt unit rated at up to 120°/s (FLIR Ranger HDC MR datasheet). Both are typical of the class, and we use them only as examples.

A small drone at useful range needs the narrow end of that zoom. Here is the arithmetic for searching a band of sky 360° wide and 20° high at the narrow end. The dwell time is an illustrative assumption, not a datasheet figure: neither datasheet gives settle or refocus times.

• Axis Q6225-LE at 2.2° × 1.3°: 164 columns × 16 rows ≈ 2,600 positions.
• Ranger HDC at 3.0° × 1.7°: 120 columns × 12 rows = 1,440 positions.
• At an assumed 0.5 s per position (slew, settle, one look), a full sweep takes about 22 minutes for the first and 12 minutes for the second. Even at 0.1 s per position, the first still takes over four minutes.
• In four minutes, a drone flying at an illustrative 30 m/s covers about 7 km.

So a zoomed PTZ cannot search. It can only look where something else tells it to. Zooming out to search fixes the coverage but gives up the magnification. At 63.8°, the camera is effectively a fixed wide-angle camera with a motor attached.

Cueing is the dependency. In practice, PTZ cameras in counter-UAS are cued by radar or radio-frequency (RF) detectors. CISA's 2025 guidance for critical infrastructure says plainly that "non-RF emitting UAS generally cannot be detected by RF systems" (CISA). Fibre-optic FPV drones, now used in large numbers in Ukraine, operate in what one Ukrainian commander called "total radio silence" (The War Zone). Radar cueing has its own failure mode. In October 2024, a Hezbollah drone that struck a base near Binyamina was tracked by Israeli radars and pursued, then dropped off the radar, likely because it flew very low (Times of Israel). A camera waiting for a cue sees nothing the cueing sensor missed.

One target at a time. A PTZ has one line of sight. Axis's own radar-cued autotracking documentation describes what happens when there are more objects than cameras: the tracker "tracks the closest object" (Axis Radar Autotracking for PTZ manual). That is a sensible design choice, and it is the swarm problem in five words. If drones arrive from opposite bearings, a pure 180° slew takes 1.2 s at 150°/s or 1.5 s at 120°/s. That is arithmetic at maximum speed, before any acceleration, settling, refocusing or re-acquisition. While the camera is pointed at the next target, the last one has no track. Third Eye Systems, a scanning-sensor vendor, makes a similar point in its own brochure, listing EO-IR challenges as "Narrow FOV, tracker logics, multiple targets" (Third Eye MeduzaX brochure, 2024).

Mechanics. Slewing, settling, wind load and vibration all come with a moving head. We have no published failure-rate data for PTZ heads in counter-UAS use, so we will only make the architectural point: when one camera is the only optical sensor covering a site, its motor and its single line of sight are a single point of failure.

Scanning panoramas: better, but still sequential

Scanning panoramic thermal imagers solve PTZ's coverage problem by rotating continuously and stitching a 360° image. They are a real step up, and vendors in this class market them for counter-UAS and swarms.

HGH's Spynel series is the best-documented example. Its 2025 leaflet describes the cameras as "scanning the panorama at 360°" with a frame rate of "up to 2 Hz". It lists 360° scan rates from 0.25 Hz to 2 Hz depending on model, panoramic resolutions up to 92,000 × 1,280 pixels, and vertical fields of view of 5°–20° for the S and X lines (HGH Spynel leaflet). HGH's drone case study recommends a 2 Hz configuration with a 20° vertical field of view, and says the system tracks "an unlimited number of targets" (HGH drone case study).

Third Eye Systems' MeduzaX pairs cooled MWIR with a daylight channel. Third Eye's brochure says it "covers 360° in seconds", is "rotated by a unique mirror technique", and lists the number of objects it can handle as "Unlimited" (Third Eye MeduzaX brochure, 2024; MeduzaX 2025 brochure). We have not found a published panoramic revisit figure beyond "in seconds", so the numbers below use HGH's published rates and labelled examples, not MeduzaX figures.

"Unlimited targets" is a software claim, and there is no reason to doubt it as a software claim. The limit is elsewhere. A scanner can hold as many tracks as its tracker allows, but it updates each track at most once per rotation. Here is what that means for a fast, crossing drone. All of this is simple arithmetic with an illustrative drone at 30 m/s, crossing at 1 km, which moves about 1.7° per second across the sensor's view.

Three consequences follow, and all three get worse with swarms.

Track association. Between two looks the tracker has to decide which new detection belongs to which old track. If two drones fly 20 m apart at 1 km, they are about 1.1° apart. At a 0.5 Hz revisit, each moves further than that between looks. With one drone, the tracker can cope. With a dozen drones in the same sector, manoeuvring, the chance of swapped or broken tracks rises. This is a standard problem in tracking: the larger the gap between updates relative to target spacing, the harder association becomes.

The fast, manoeuvring drone. You do not need a swarm to expose the gap; one fast drone that turns is enough. A tracker predicts where a target will be at the next look, usually by assuming it keeps its course and speed. Take an illustrative drone at 30 m/s, 300 m from the sensor, crossing the line of sight: it moves about 5.7° per second across the sky. At a 0.5 Hz revisit that is about 11° between looks, and at 2 Hz about 3°. If it turns or dives in between, it reappears far from where the prediction said, and the tracker has to decide whether that is the same drone, a new one or noise. A staring sensor at 30 frames/s sees the same drone move about 0.2° per frame and watches the turn happen. The faster and closer the drone, and the harder it manoeuvres, the bigger the gap. Those are the drones that matter most in the last seconds before they reach the site.

Time on target. A scanner sees a given bearing only for the brief moment the strip sweeps past it, once per rotation. A drone that manoeuvres, drops behind clutter or dives in the last seconds has fewer chances to be seen in that period.

Confirmation versus delay. A detection system usually confirms a target by seeing it consistently over several looks. Single-look noise, birds and clutter tend not to persist, and drones do. This is our inference, not a vendor figure. A sensor with fewer looks per second must either confirm on fewer looks, which risks more false alarms, or wait for more looks, which adds delay. A staring sensor gets the same evidence in a fraction of the time. We are not claiming that any named scanner has a high false-alarm rate. Third Eye publishes its own low false-alarm figures, which we have not tested. The point is that a lower update rate narrows the margin between false alarms and latency.

Scanners can trade coverage for speed. HGH's leaflet lists higher sector scan rates, for example 2.5 Hz on a 90° sector for one Spynel-M model. That helps, but it gives up the full panorama, which a multi-bearing swarm needs.

Where these sensors are genuinely the right tool

None of this makes PTZ or scanners obsolete.

A PTZ is the right tool for identification after detection. Zooming onto a confirmed track gives an operator the detail to classify the airframe, see the payload, capture evidence and decide on a response. Axis itself sells exactly that pairing: a fixed thermal camera running analytics steers the PTZ, so you get close-ups "without compromising detection coverage since the fixed camera maintains vigilance" (Axis Perimeter Defender PTZ Autotracking). That is the architecture this post recommends: fixed sensors detect, and the PTZ identifies.

A scanning panorama has an advantage a staring sensor at the same price cannot match: angular resolution. Spread 92,000 pixels across 360° and you get about 256 pixels per degree. That translates into longer detection range per pixel, and it matters for early warning against slower targets at distance.

HGH's own case study makes the broader point well: radar, PTZ and RF have "all disappointed, when used as a single technology" (HGH). CISA reaches the same conclusion from the buyer's side, recommending a "system of systems" with sensor fusion (CISA).

The alternative: persistent staring coverage

The approach we build is to cover the sky with fixed, overlapping, wide-field thermal cameras, each watching its own sector in every frame. Our nodes use FLIR Boson thermal cores. With the 14 mm lens, the Boson 640 has a 32° horizontal field of view at 640 × 512 and outputs at up to 60 Hz (Teledyne FLIR Boson 640, 14 mm). Detection and tracking run on the node, which reports tracks to a hub and over SAPIENT (BSI Flex 335). The UK MOD adopted SAPIENT as its counter-UAS standard (UK Government).

What staring gives you:

• No unwatched direction. Every covered bearing is imaged every frame, so the revisit interval is one frame period.
• Many simultaneous tracks at frame rate. Each camera tracks everything in its sector, updated every frame, so a drone moves about a pixel between updates, not dozens.
• No dependency on cooperation. A thermal camera sees a drone by its heat and shape. That covers fibre-optic and pre-programmed drones as well as radio-controlled ones. CISA notes that EO/IR sensors "can detect non-RF emitting UAS".
• Graceful failure. Losing one camera leaves one sector uncovered, not the whole site, and nothing in the optical path moves.
• A clean handoff to identification. A staring track is the cue a PTZ needs, and SAPIENT carries it to whatever fusion node and effector sit behind it.

What staring costs, honestly

Staring has real trade-offs.

It takes many cameras. At 32° per camera, a full 360° ring needs at least 12 cameras with no overlap, and more with overlap or a taller vertical band. That is more units to site, power and maintain than one rotating head.

Lower angular resolution, so shorter range per pixel. 640 pixels across 32° is about 20 pixels per degree. The Spynel models above range from about 32 to 256 pixels per degree. At the same detector size, a staring camera sees a given drone across fewer pixels, so it detects it later. Against a slow, distant scout, a high-resolution scanner can see it first.

Uncooled cores are less sensitive than cooled MWIR. That is the price of small, low-power nodes.

It does not identify by itself at range. A few pixels can say "drone-like object, here, moving like this". They cannot always say which drone. That is why the zoom camera belongs behind the staring layer.

Whether staring coverage is the better choice for a given site depends on the threat. Against fast, many-bearing, swarm-style attacks inside a few kilometres, we think persistence matters more than per-pixel range. Against a single slow target at long range, the trade can go the other way.

The takeaway

PTZ and scanning panoramas share a design choice: they look at the sky one piece at a time. That works when targets are few, slow or cued. It struggles when many targets arrive at once, from several directions, quickly and without emitting.

Detect with sensors that never look away. Identify with sensors that can zoom in. Put both on one common picture, because no single sensor wins the drone war.

Independent commentary by GoTeam. Not affiliated with or endorsed by any organisation named above. Product specifications are quoted from each manufacturer's own published material and have not been independently tested. Figures labelled illustrative or arithmetic are ours, not the manufacturers'. Our reading of how sequential sensors behave against swarms is analysis, not a measured comparison of the named products.