What AI video analytics can and can't do for a monitored site

"AI-powered" is on the box of nearly every camera and monitoring product sold today. Some of it is transformative. Some of it is motion detection with a new name. If you're choosing a system, or deciding what to promise a client, here is a plain account of what the technology does well and where it stops.

What it does well

Telling people and vehicles from everything else. This is the big one. Modern object detection reliably separates a person or a car from rain, foliage, animals and changing light, which is where most false alarms come from.

Applying rules consistently, at 3am, on every camera. A person crossing a line, a vehicle stopping in a no-waiting zone, someone lingering at a fence. Software watches every feed all the time and doesn't get tired at 4am.

Finding things in recorded footage. "Every white van at the gate last Tuesday" takes seconds when footage has been indexed, not hours of scrubbing.

Reading number plates at entrances and exits, where the camera is set up for it.

Counting. People through a door, vehicles through a gate, occupancy of a yard over a day.

What it does less well

Judging intent. Software can see that someone is by the fence at 2am. It can't know whether they're a trespasser or the site manager who forgot their keys. It can flag the event. A person decides what it means.

Poor footage. A camera with a dirty lens, pointed into the sun or showing a figure twelve pixels high won't become reliable because the software is clever. Analytics are only as good as the camera's view.

Unusual scenes. Models are trained on the kinds of things they usually see. An odd camera angle, a scene full of mannequins, or heavy fog will produce more errors than a clear view of a car park.

Everything, everywhere, out of the box. Each site needs its zones, schedules and thresholds set. The first week always involves some tuning.

Where people are essential

The strongest setups use analytics to filter and prioritise, and people to verify and respond. Software reduces thousands of motion events to a handful of real ones, and an operator looks at each with fresh eyes and the context to act. That's a far better use of skilled people than dismissing foxes all night.

Questions to ask a vendor

• Does it work with the cameras I already have, or do I need new ones?
• Where does the analysis run: on site, in the cloud, or both? What happens when the internet drops?
• How do I tune a noisy camera, and can I see which cameras cause the most false alarms?
• Can I see what the system saw, the clip and the detection, behind every alert?
• What happens to the footage, and how long is it kept?

The best answer to "can AI do this?" is usually "yes, with the right camera view and a person looking at what it flags". Be wary of anyone who leaves out the second half.