Drawing Perimeter Zones That Survive Palm Shadow and Blowing Dust

Perimeter analytics answers a narrow question: did something enter this area, at a time when nothing should have. Both halves of that are configuration — the area and the time — and almost all the noise a site complains about traces back to one of them being wrong rather than to the model being poor.

Every false alarm has a physical cause. Below is the catalogue for outdoor sites in Oman, and what defeats each one.

The catalogue of causes

Read the right-hand column and a pattern appears: almost nothing on that list is fixed by adjusting sensitivity. Turning sensitivity down to stop the cats is what makes the system miss a person, which is the exact trade every frustrated site makes at about week three.

The single most common mistake: drawing the fence

The instinct is to draw the detection zone along the boundary itself. This is wrong in both directions. It catches everything happening on the public side — pedestrians on the pavement, cars parking, a delivery van reversing — none of which is an intrusion. And it gives you no warning time, because by the time something is detected it is already at the fence.

Draw it a few metres inside instead. Anything in that band has already crossed, which is unambiguous, and you have gained the seconds it takes to cross open ground. The band should also be wide enough that a person walking normally appears in several consecutive frames, because a zone too thin to dwell in is a zone that misses fast movement.

Schedules are half the system and get a tenth of the attention

The same person crossing the same yard is routine at 07:30 and an incident at 02:30. No amount of visual analysis establishes that; only a schedule does. Sites that run one rule around the clock generate hundreds of daytime alerts nobody reads, and the habit of not reading alerts is what makes the 02:30 one useless too.

Schedules need to reflect how the site actually runs, which means Friday, public holidays, Ramadan hours and the shutdown week are all separate cases rather than exceptions someone remembers to configure. A practical starting point is three regimes: open hours with detection logging but not alerting, closed hours with alerting, and a hardened overnight window with escalation. That distinction between logging and alerting is the one most systems collapse and the one that most reduces noise.

Two configuration rules that outperform tuning

Require dwell. Insisting an object remains in the zone for a short interval before alerting removes almost all of the transient causes above — a shadow crossing, a bird, a light sweep — while barely affecting a real intruder, who has to spend time in the space to get anywhere. A second or two of required dwell is usually worth more than any threshold change.

Filter by object class and size. Alerting only on person and vehicle, with a minimum pixel size floor, eliminates the animal problem without reducing sensitivity to people. The size floor matters as much as the class: it prevents a cat close to the lens from being classified as a distant person, which is the failure that makes class filtering look unreliable.