People ask us this in two very different tones. Some want reassurance that it isn't magic. Others want reassurance that it is.

It's neither. Here's what actually happens.

Motion detection: what you probably have now

Conventional CCTV motion detection compares one video frame to the next and measures how many pixels changed. Past a threshold, it fires.

That's the whole mechanism. It has no idea what changed — only that something did.

This is why your current system alerts for rain, swaying branches, shadows crossing a wall, headlights sweeping a compound, and insects landing on the lens at night. All of those change pixels. From the DVR's point of view they are indistinguishable from a person climbing your fence.

AI intruder detection: classify first, then decide

An AI system inverts the order of operations. Instead of asking "did something change", it asks "what is in this frame".

A detection model runs over each frame and returns objects with labels and positions: person at these coordinates, vehicle at these, dog at these. It does this many times a second, and it tracks each object across frames, so it knows that the person in frame 40 is the same person from frame 12.

Only after that does any alerting logic run.

The three questions before your phone buzzes

Once the system knows what it's looking at, an alert has to pass three filters:

Is it a class I care about? You almost certainly want people. You may want vehicles. You almost certainly do not want animals — stray dogs at a factory perimeter are the single biggest source of night-time false alarms in India.

Is it in a zone I marked? You draw a line or a region on the camera view — the compound wall, the loading bay, the stock room door. Movement outside that zone is recorded but never interrupts anybody.

Is it during hours I armed? The dispatch bay is busy until 8pm and should be silent after. The pharmacy store should be alarmed around the clock. Each zone carries its own schedule.

A person in an armed zone during armed hours produces a notification with a snapshot, typically under three seconds from the moment they crossed the line. Everything else is logged and left alone.

Where the identity layer comes in

If faces are enrolled, there's a fourth filter: is this someone I know?

Your night supervisor walking through the yard at 2am is a person, in an armed zone, during armed hours — and still not something you want to be woken for. Recognition lets the system stay silent for known people while remaining uncompromising about unknown ones.

This is also what makes daytime alerting survivable at any site with staff.

Where it runs, and why that matters

All of this runs on a small edge device at your site rather than in a cloud. Two consequences follow, and both matter more than they sound:

Your video stays on your premises. Nothing is uploaded to be analysed, so the privacy question has a straightforward answer.

And detection survives your internet going down. The connection is only needed to deliver notifications; if it drops, detection and any local response — siren, strobe, floodlight — carry on regardless.

What it doesn't do

It doesn't know intent. It knows a person is somewhere they shouldn't be at a time they shouldn't be there, which is a strong signal and not a verdict. Every alert deserves a human glance before anyone acts on it.

It also doesn't identify strangers. It can tell you this is not anyone enrolled — it cannot tell you their name. Any vendor implying otherwise is selling you something that doesn't exist.

The honest summary

AI intruder detection is object classification plus zones plus schedules plus, optionally, identity. That's it. There's no intuition and no learning about your habits in the way the marketing sometimes implies.

What makes it worth having is not sophistication. It's that the resulting alerts are rare enough and accurate enough that you leave notifications switched on — which is the only state in which any security system does anything at all.

How does AI intruder detection actually work?