Radar and EO/IR sensor fusion: how combined surveillance works

An electro-optical / infrared (EO/IR) unit combines a daylight camera with a thermal imager, usually on a pan-tilt head with a zoom lens. It adds three things radar does not give you.

Visual and thermal identity. Radar sees a reflection; it has little to say about shape, markings, payload or whether a small contact is a bird, a drone or something else. An image does, and a thermal image shows heat signatures as well. This is the evidence a human operator trusts when deciding what a contact is.

Day and night. The thermal channel works in darkness, so day-night surveillance does not depend on ambient light. The daylight channel adds detail when light is good.

Passive confirmation. A camera does not transmit. It observes without emitting anything an adversary can detect or locate, which matters in sensitive environments and is one reason EO/IR is a favourite second sensor beside an emitting radar.

A camera has a narrow field of view: it sees only where it points, so on its own it must search slowly. Radar covers the opposite ground. It scans a wide area continuously and keeps working in cloud, rain, haze and darkness, where optical and thermal performance can degrade.

Radar also measures — range, bearing and radial velocity of each contact — and keeps a track over time. A single passive camera gives a good bearing and angular size but struggles to give range on its own. So the radar supplies the where and the how fast; the camera supplies the what.

In practice the pairing runs as a short loop. 1. Detect: the radar picks up a contact and starts a track. 2. Cue: the track's position and motion are converted into pointing angles and sent to the EO/IR head, which slews to the contact — slew-to-cue — and sets zoom and focus to suit the estimated range. 3. Confirm: the camera acquires the contact and the image is checked, by an operator, by a classifier, or both, to establish what it is. 4. Fuse: the camera detection and the radar track are tied to one fused track carrying position, velocity and identity evidence.

Cueing is what makes the camera useful at area scale. Without it, an operator hunts across a wide sky with a narrow lens; with it, the camera arrives where the radar already knows to look, and the operator's attention goes to judging the image rather than finding the contact.

Fusion only works if the two sensors agree about where and when. Space: the radar and the camera sit in different places with different reference frames, so their positions must be registered into a common frame, and the camera's pointing must be calibrated against the radar's measurements. Time: the radar updates at its scan rate, the camera at its frame rate, and a moving contact is somewhere different at each instant, so measurements are time-stamped and extrapolated to a common moment before they are compared.

Even then the sensors will sometimes disagree. The radar reports a contact the camera cannot see — it may be beyond the lens, behind cloud, or simply not visible. The camera shows an object the radar does not report, perhaps too small or slow to stand out from clutter. Or two contacts sit inside one camera frame. Good fusion does not pretend these away: it keeps the disagreement visible, attaches a confidence to each track, and flags contacts that remain unconfirmed rather than quietly promoting or discarding them. This is the association problem described in the multi-sensor fusion guide.

Automatic cueing. Software can drive the camera from radar tracks without a person steering it, prioritising which contact to look at next when several compete for one lens.

Visual classification. A trained model can recognise the class of what the camera sees — for example separating a small drone from a bird or other airborne clutter — and report a confidence with its reasons, giving the operator a head start on the judgement.

Holding the fused track. Learned association keeps radar and camera detections attached to the same object as it manoeuvres, passes behind obstacles, or leaves the camera's view for a few seconds, instead of splitting one contact into two ghost tracks.

Camera frames and radar tracks are sensitive. Running cueing, classification and fusion at the edge, inside the operator's own perimeter, avoids sending surveillance imagery to a third-party cloud, and keeps the picture available when links are degraded. Our wider view of why this matters is in the air-surveillance pillar.

AgentAnywhere's posture here is decision support: the AI cues, classifies and recommends with its reasons, a human makes every consequential call, and each recommendation carries an inspectable record. Sanjaya is our sovereign edge-AI family for sensor telemetry and anomaly detection; see the sovereign model portfolio.

AI in radar & air surveillance — the pillar: how AI turns raw returns into a recognised air picture.

Radar signal processing — what happens to the echo — detection, Doppler, CFAR and tracking.

Clutter, noise & false targets — why radar sees things that aren't there, and how it is cleaned up.

Multi-sensor fusion — turning radar, IFF and EO/IR that disagree into one trusted track.

AI in signal processing — learned detection, recognition and ECCM beyond fixed thresholds.

Low-observable & low-flying targets — why small, slow, low drones are hard — and how AI pulls them out.