The 4 Levels of
AI Security Cameras

A warehouse manager in Birmingham gets a call: someone walked off with £10,000 of stock. He has 14 cameras and 30 days of footage. His team spends four hours scrubbing through it just to find 40 seconds of useful video.

That's not a security system. That's an expensive hard drive. The cameras saw everything and told him nothing — which is the gap AI actually closes. Below are the four levels of what's possible with the cameras you already own, from smart recording all the way to a system you can ask questions.

Prefer to watch? The full walkthrough is above. This post goes deeper on the tools and the trade-offs at each level.

Level 1 is an NVR — a network video recorder. A computer sitting on your network, quietly recording the streams from every camera you care about, so the footage is ready when you need it. The Birmingham warehouse already had this. Most businesses do. The interesting question is what you use to detect something.

The simplest answer is motion detection. The most basic version compares two consecutive frames and flags any difference. The problem: it flags everything — a curtain in a draught, a tree, a fan spinning in the corner. In the warehouse, that's roughly 10,000 detections a day. Useless.

That sequence-based approach is the upgrade. It can't tell a person from a stray cat, but it kills the constant-motion noise — bringing the warehouse from 10,000 down to about 2,000 detections a day. Those are your level-1 alerts.

Level 2 is real-time alerts, and the technology that unlocks it is object detection. You hand the model an image and it tells you what's in it and where: "three people, here, here and here" — each wrapped in a bounding box. Now an alert can mean something specific instead of "something moved."

What a model can detect depends entirely on what it was trained on. Train on the COCO dataset and you get 80 everyday classes — person, car, dog, bag. For trespassing in a forbidden zone, that's already enough: detect a person where one shouldn't be, and you're done.

Harder cases need their own data. For Angel Protection, we built firearm detection for US schools under two brutal constraints: a top-down CCTV angle, and a gun that appears only a few dozen pixels wide. Off-the-shelf datasets had neither — so we built one. Here's how that system works across 1,000+ cameras →

Levels 1 and 2 look at single frames. Level 3 looks at sequences — analysing the video, not the image. That shift unlocks an entirely different class of question: not "is there a person?" but "what is that person doing?"

Tracking (DeepSORT and friends) follows each person across frames — perfect for loitering and store flow analysis. Pose estimation maps the body's joints, and feeding that to a classifier catches fights, violence, or — more usefully — a person falling. Detect more than one object class at once and "person + bag, then person leaves" becomes unattended-luggage detection. The hardware bump over level 2 is real but modest.

Level 4 is conversational. You query your footage the same way you'd ask ChatGPT or Claude a question — in plain language. The interesting things you want to know are rarely the ones you set up a detector for in advance, and this is what finally lets you go looking for them.

The catch: this takes a lot more horsepower. Even the smaller vision-language models need far beefier hardware than anything at levels 1–3, which means real money. The pay-off isn't a smaller alert count — it's open-ended exploration. Instead of waiting for a theft, you can ask: "is anyone behaving like they're casing the place?" and go looking before it happens.