Why motion alerts cry wolf — and what actually fixes it

Nearly everyone who switches on CCTV notifications switches them off again within two weeks. The reason is not a bad camera — it is what "motion detection" actually measures.

What motion detection really does

Classic motion detection does not see objects. It compares one video frame with the next and measures how many pixels changed. If enough pixels change in a region, it declares "motion". That is the entire mechanism.

This means it fires on anything that changes pixels:

• A moth or spider on the lens (huge pixel change, right in front of the camera)
• Trees and washing moving in wind
• Rain, and especially insects lit by infrared at night
• Car headlights sweeping across a wall
• Clouds moving across the sun, changing the whole scene's brightness
• The camera's own infrared switching on at dusk

None of these are threats, and all of them look identical to a person walking, because the algorithm has no idea what a person is.

Why turning down sensitivity doesn't work

The usual advice is to reduce sensitivity. The trouble is that sensitivity is a single dial trading false alarms against missed events. Turn it down far enough to stop alerting on a cat, and a person walking at the far end of the driveway — smaller in frame than the nearby cat — stops registering too. You have not solved the problem; you have moved it somewhere more dangerous.

The three things that genuinely reduce false alarms

1. Detect objects, not motion

Modern object detection runs a neural network over the frame and returns what it found, with a confidence score: person 0.91, cat 0.78, car 0.85. Because it recognises the thing rather than the change, wind and rain produce nothing at all. You then choose what you care about — usually people, sometimes vehicles.

This is the single biggest improvement available, and it is why "AI camera" systems feel so different in daily use.

2. Draw the zone properly

Even perfect person-detection will alert on every pedestrian if the camera can see the pavement. On a gate camera in a South African suburb that might be forty people a day, all correctly detected, none of them your problem.

Restricting detection to the area inside your wall — the driveway, the path to the door — is what converts a technically accurate system into a useful one. In our own testing this single change accounted for the majority of unwanted alerts on a street-facing camera.

3. Require persistence

Set a minimum number of frames an object must be present before it counts. Something that clips the corner of a zone for a fraction of a second is usually a bird or a passing car; a person walking in is present for seconds. A small persistence requirement removes a whole class of edge-of-zone noise.

The setting nobody mentions: what happens after too many alerts

Even a well-tuned system has bad days — a delivery, a gardener, a party. If it keeps shouting, you will start ignoring it, and the tuning was pointless.

A good system notices its own noise: if one camera has alerted repeatedly in a day, it should rest that camera for a while, tell you it has done so and why, and then resume automatically. That protects the thing that actually matters — your willingness to look at the next alert.

What "good" looks like

Symptom	Cause	Fix
Alerts at night in rain	IR-lit insects and droplets	Object detection
Alerts every few minutes daytime	Zone includes street/pavement	Redraw the zone
Alerts on cats and dogs	Motion, or a model that confuses small animals	Person-only detection, higher confidence
Brief alerts with nothing in frame	Object clipping the zone edge	Require persistence
You ignore all of them	Alert fatigue	Auto-rest a noisy camera; escalate real ones by phone