Technology

Radar, LiDAR, thermal and artificial intelligence

Explained without marketing: what each technology does, what it does not do, and when it is worth paying for. Because a radar in a small garden is money badly spent.

Radar

It sees movement, not light

FMCW radar: simultaneous coverage and object trackingPlan view of the area covered by a security radar. The radar does not rotate: it observes the whole sector at once using digital beamforming, and each measurement cycle yields range, azimuth and velocity for every object, which it follows as a track. 20 m40 m60 m-90°-60°-30°+0°+30°+60°+90° WALL RADAR SHADOW PERSON · ID 04 v = 1.4 m/s VEHICLE · ID 05 'SWAYING OBJECTS' FILTER vegetation discarded < 3 m: two objects merge into one 24 GHz FMCW RADAR NO MOVING PARTS · FULL SECTOR EVERY CYCLE AZIMUTH 180° · RANGE 60 m
Radar. There is no rotating antenna: the whole sector is measured every cycle. Each object yields range, azimuth and velocity, and from those a track is built. Classification separates person and vehicle; everything else — including a large dog or a wild boar — stays 'unknown'. Behind every solid obstacle there is a blind zone, and two objects less than 3 m apart merge into a single detection.

Radar works like the one at an airport, applied to your plot: it doesn't interpret an image, it detects moving mass. That's why shadows, car headlights, rain and an insect sitting on the lens don't affect it — and those are exactly the four classic causes of a false alarm on a camera.

Professional units can be told to ignore small objects — the manufacturer's own documentation specifically mentions cats and rabbits — and to ignore swaying objects, meaning trees and bushes. In a Mediterranean garden with olive trees, oleander and palms, that second function alone justifies radar over a motion detector.

When it makes sense

  • Open, unobstructed ground, roughly 500 to 20,000 m²
  • Plenty of vegetation that moves in the wind
  • A long driveway, with early warning of an approaching vehicle

And what radar does not do.

It cannot see through solid objects: a wall, a large hedge, a retaining wall or the pool plant room all create a radar shadow. In a heavily subdivided garden you would need more than one unit, and that pushes the price up sharply: we will tell you that beforehand, not afterwards.

And it cannot tell a large dog from a person: it classifies it as "unknown". That is why radar is always installed alongside an AI camera.

Does it emit radiation? The transmitted power is under 100 milliwatts — a fraction of a mobile phone's — and it operates in a licence-free band across the whole European Union.

LiDAR

An invisible laser curtain

2D LiDAR: range profile and detection by deviationPlan view. A rotating mirror sweeps the plane and measures range at every angular step. The unit learns the background profile; a deviation from that profile is a detection, and object size is computed from its angular width and its range. 123 DEVIATION = DETECTION width 62 cm · DETECTED width 26 cm · BELOW THRESHOLD 2D LiDAR · CLASS 1 ROTATING MIRROR WALL / FENCE the mirror turns 360°; the useful sector is 95° LEARNED BACKGROUND PROFILE MEASURED PROFILE OBJECT-SIZE THRESHOLD PER ZONE ZONE 1 · ≥ 60 cm ZONE 2 · ≥ 30 cm ZONE 3 · ≥ 15 cm size = angular width × range
LiDAR. The mirror rotates continuously and measures a range at every angular step: the result is a range profile. The unit learns the background profile, and what it detects is the deviation from it. Because it knows both range and angular width, it computes the real size of the object — which is why the threshold can be set in centimetres, zone by zone.

LiDAR projects an invisible plane of laser light and measures the distance at which it is broken. Its distinguishing feature is unique on the market: you can define, zone by zone, what size of object counts. It is the only technical way to reliably ignore a 30 cm cat and still detect a person crawling along the ground. No motion detector can do that.

With it you build, quite literally, a wall of light in front of a specific surface.

Where we install it

  • In front of a glazed living room or a run of terrace doors
  • Wine cellar, safe room or a wall holding artwork
  • Along the top of a wall: a virtual wall above the real one
  • A narrow strip along the boundary, where radar would see the neighbour's plot

When we don't recommend it.

For general outdoor surveillance it is expensive, and radar does that job better for less. LiDAR is only justified when there is one specific critical plane to protect.

In dense fog or very heavy rain its performance degrades as a matter of physics: scattering absorbs the pulse. Filtering modes reduce that, they don't remove it.

Is it dangerous to the eyes? No. The units we install are Class 1, the lowest laser classification — the same category as a supermarket barcode scanner. No warning signage is required.

Thermal

Heat instead of image

Thermal: pixels on target at three distancesThe same person seen by a 160 by 120 thermal camera at 91, 23 and 11 metres. Angular resolution fixes how many sensor pixels the target occupies: at 91 metres only a few, enough to say a heat source is present; at 23 metres a human figure is recognisable; at 11 metres features can be told apart. SENSOR 160 × 120 · NETD < 40 mK · 3.1 mm LENS 91 m · DETECTION ≈ 3 px on target a heat source is present 23 m · RECOGNITION ≈ 9 px on target it is a person 11 m · IDENTIFICATION ≈ 18 px on target features distinguishable Angular resolution does not change: what changes is how many pixels the target covers.
Thermal. Distance does not 'blur' the image: it reduces the number of sensor pixels the target occupies. That is where the three catalogue figures come from. The detection figure is the one that gets advertised; the recognition figure is the one that decides whether the system is useful, and it is the one we size from.

A thermal camera doesn't see a picture: it sees heat. It works in absolute darkness, can't be blinded with a torch, and isn't fooled by shadows or reflections. On a property with no lighting — and no intention of installing any — it is the only genuinely reliable night-time detection.

The figure almost nobody will tell you. A thermal camera detects much further than it recognises. A typical model notices that someone is there at about 90 metres, but only confirms it is a person at about 23 metres. If a quote argues using the 90-metre figure, you are being sold the wrong number. We design using the recognition figure.

When it is needed

  • Long sight lines, beyond roughly 60–80 metres
  • Unlit areas with no possibility of installing lighting
  • Rural plots and inland country properties

When it isn't

  • If the area can be lit and an AI camera gives colour images at night
  • On small plots: cost without return

We install bi-spectral units: thermal and visible channel in one housing. The thermal side raises the alert; the visible channel provides the image that serves as evidence.

Artificial intelligence

The camera that knows what it's looking at

There are three levels, and the difference between them decides whether you sleep well.

Classifier output: bounding box, class, confidence and trackingFrame from a camera with analytics. Every detected object produces a bounding box, a class, a confidence value compared against a configurable threshold, and a tracking identifier. A person raises an alarm; a vehicle is logged; an animal is logged without raising an alarm. PERSON · ID 04 ALARM VEHICLE · ID 05 LOG ANIMAL · ID 06 LOG · NO ALARM AI CAMERA · ON-DEVICE INFERENCE EMITTED METADATA classperson / vehicle / animal confidencevalue vs. configurable threshold boxx, y, width, height trackingpersistent identifier actionalarm / log only confidence threshold
AI analytics. The detector does not say 'movement': for every object it emits class, confidence, bounding box and a tracking id. The action depends on the class and on whether confidence clears the configured threshold. The animal category is what lets a wild boar be recorded without waking anyone.
LEVEL 1

Basic filtering

Distinguishes person and vehicle from generic movement. Removes leaves, shadows and light changes. This is today's acceptable minimum.

LEVEL 2

On-camera classification

Person, vehicle and vehicle type, processed on the device itself, with no server and no cloud.

LEVEL 3 · THE ONE THAT MATTERS HERE

Animal category

Person, vehicle and animal as a category in its own right. The "animal" event is recorded, but it doesn't wake you at night.

Why the animal category matters particularly in Malaga. A study by the Malaga Veterinary Association together with the University of Cordoba analysed urban wild boar captured in the province: 40% in Marbella, 30% in Malaga city, 22.5% in Fuengirola and 7.5% in Benahavis. A wild boar weighs many times more than the 25 kg "pet immunity" that detectors declare, and it moves at ground level — exactly where the sensor is looking. Neither a motion detector nor a radar solves it. Only image classification does.

And if you already have cameras. They don't always have to be replaced: an analysis unit can add person, vehicle and animal recognition to existing ONVIF cameras. It is a cheaper way in, and it can be extended later.

Summary

What to choose for your plot

SituationWhat makes sense
Small, open plot up to ~500 m²AI cameras and detection at the fence. No radar: that would be over-specifying.
Large garden with vegetation and windRadar plus an AI camera.
Glazed frontage, cellar or a wall of valueLiDAR as a critical plane, in addition to the above.
Large property, long sight lines, no lightingRadar, bi-spectral thermal and mandatory AI classification.
Large dog or wild boar in the areaAI classification with an animal category. Essential.