Perimeter intrusion alerts: from detection to video review

Perimeter detection has the worst signal-to-noise ratio of anything in physical security. A fence line reports wind, foliage, rain, cats, foxes and the occasional person, and reports them all identically. The result is predictable: within a few weeks, nobody acts on any of them.

Adding a camera view to the alert is what breaks that cycle — but only if the camera can actually see the thing.

The chain, end to end

1. Something crosses the line. Either a sensor fires, or — more usefully — a camera watching a marked area detects a person inside it during the hours it should be empty.

2. The event carries a frame. The image at the moment of detection, captured as part of normal operation rather than fetched from a recorder afterwards.

3. A person decides. Looking at the frame, the question “is this worth a response” takes about two seconds. Without the frame it takes a phone call and a drive.

4. The area responds. A floodlight, a voice challenge from a speaker covering that spot, a siren where the site permits it. This is where perimeter detection pays for itself, because it happens while the person is still standing there.

5. The evidence stays. Frame, clip, camera, area and time, held together for the investigation or the insurance claim.

Zones and schedules beat thresholds

The instinct when a boundary produces too many alerts is to turn the sensitivity down. It is almost always the wrong move: it suppresses genuine detections everywhere while barely touching the specific nuisance.

What works instead:

Draw the area tightly. Not the whole camera view — the strip of ground that matters. A camera watching its entire field reports everything in it, including the public road at the top of frame.

Give each area its own hours. A yard that is busy until 19:00 and should be empty afterwards is a different problem before and after that time, and saying so removes most of the daytime noise for free.

Look for a person, not for movement. Motion detection fires on branches. Person detection does not. This single distinction removes the bulk of weather-driven noise.

Where it genuinely fails

This is the part worth reading twice, because it decides whether a perimeter project works.

Darkness. If the area is unlit and the camera has no usable infrared at that distance, there is nothing in the image to detect. No model fixes an image that does not contain the information. The fix is a light — which also improves the evidence and deters the person.

Distance and lens. A person thirty metres away at the edge of a wide-angle view occupies very few pixels. The same camera at ten metres is a different proposition entirely. What matters is pixels-on-target at the distance where events happen, not the megapixel number on the datasheet.

Lens contamination. Spiders, condensation and dust at the lens are a leading cause of night-time nuisance events on boundary cameras. Cleaning schedules are an unglamorous but real part of making this work.

Camera placement. A camera mounted high and looking down a fence gives you a long, thin, foreshortened view. One looking across the approach gives you a person-shaped object. Same hardware, completely different result.

Three of those four are hardware and siting problems. That is why a survey that walks the boundary is worth more than any amount of configuration afterwards.

What to do with existing perimeter sensors

If you already run fence sensors, beams or radar, keep them. They are not in competition with a camera-based approach — they are a different trigger for the same question. The useful work is mapping which camera covers each sensor zone so that a sensor event and a camera view end up in the same place.

What the camera adds is the answer to what crossed the line, which is the question the sensor was never able to answer.

What a proof of concept should measure

Run it for at least two weeks, across real weather, and record:

• Events per night, per area. The absolute number, before any tuning.
• How many were a person. The ratio is your real signal-to-noise figure.
• Whether the night image is usable on each boundary camera, judged by looking at it.
• Time from detection to the on-call person having the frame.
• What the response should be per area — light, voice, siren, silent — which is usually different per area and worth deciding deliberately.

Two weeks of your own numbers will tell you more than any published benchmark, because the benchmark was measured on a boundary that is not yours.

The honest summary

Perimeter detection works well when the camera can see, the area is drawn tightly, the hours are set, and something happens in the area when an event fires. It works poorly in the dark at distance, and no software setting changes that.

Anyone who tells you otherwise has not walked your fence line.