Fundamentals
What is AI video analytics? A plain explanation
Software that watches a video feed and works out what is in it — a person, a vehicle, a piece of equipment — and whether what they are doing matters enough to tell someone about.
VuePoint
AI video analytics is software that interprets a video feed automatically. It identifies what is in the frame — a person, a vehicle, a forklift, a hard hat — tracks how those things move, and decides whether the combination is worth telling someone about. The output is not video. It is a short list of events, each with a time, a place, and a clip attached.
The term intelligent video analytics means the same thing. So does AI CCTV in British and Australian usage. All three describe the same shift: from a system that records everything and expects a person to find the important part, to one that identifies the important part and hands it over.
What problem it actually solves
Cameras stopped being the constraint a long time ago. Storage is cheap, sensors are good, and most sites that want coverage have it. The constraint is that recorded video is only useful if somebody looks at it, and nobody looks at it.
The realistic pattern on most sites is that footage is reviewed after something has already gone wrong, by which point nobody knows which of the last four days to search. The recording existed the whole time. The knowledge did not.
Motion detection was the first attempt at fixing this, and it fails for a specific reason: it detects change in pixels, not events. Rain, headlights sweeping a wall, a flag, a spider on the lens, and a tree in wind all produce motion. So does a person climbing a fence. A system that cannot tell those apart sends alerts for all of them, and a system that alerts constantly gets muted within a fortnight.
How it works, without the jargon
The four stages
- 1
Detection — what is in the frame
A model trained on very large numbers of labelled images draws a box around each thing it recognises and names it: person, vehicle, bicycle, animal. This is the part that separates a person from a shadow, and it is why a modern system does not alert on rain.
- 2
Tracking — where it goes
Each detected object is followed frame to frame, so the software knows it is one person crossing a yard rather than thirty separate detections. Tracking is what makes duration and direction available at all.
- 3
Rules — whether it matters
Detection plus tracking produces facts. A rule turns facts into an event: a person inside this boundary, after these hours, for longer than this many seconds. The rule is where site knowledge lives, and it is the part most often set up badly.
- 4
Delivery — telling a person
The event goes to whoever is responsible, with a clip and a timestamp. Everything else stays unremarkable. The measure of a good system is how little it sends, not how much.
What it can realistically detect
| Detection | How reliable | What it depends on |
|---|---|---|
| Person present in an area | Very reliable | The core capability. Works well in most conditions with adequate light or IR |
| Vehicle present, and its type | Very reliable | Large, high-contrast, predictable shapes |
| Crossing a line or entering a zone | Reliable | Requires the boundary to be drawn where the camera can actually see it |
| Loitering — presence beyond a duration | Reliable | Tracking has to hold through partial occlusion |
| Object left behind or removed | Reliable with persistence checks | Confirming the change is still there minutes later removes most false positives |
| PPE present or absent — hard hat, hi-vis | Good at close range | Distance and angle. A wide site-overview camera will not do this |
| Licence plate capture | Placement-dependent | A dedicated camera at a shallow angle on the approach. Not a software question |
| Identifying who a person is | Not what this is | That is facial recognition — a separate technology, separately regulated, and a different product |
Edge or cloud — and why it matters more than it sounds
The analysis can run on hardware at the camera (at the edge) or on servers elsewhere (in the cloud). On a site with fibre this is an architecture preference. On a site without it, it decides whether the system is possible at all.
Continuous video from several cameras over a cellular link will exhaust a data plan in days and cost more than the cameras. Running the analysis on the unit and sending only an event and a still image is a different order of magnitude of data. This is why cameras on sites with no power and no network are an edge problem before they are anything else.
Do you need new cameras?
Usually not. Analytics is software, and it can run against cameras that are already installed and already recording — which is often the cheapest place to start, because the footage is already being captured and simply is not surfacing anything.
New hardware earns its place when there is no power or network where the coverage is needed, or when an existing camera is pointed somewhere useless for the event you care about. A camera framed for a general view of a yard will not capture a licence plate, and no software fixes that. Placement is the decision that determines whether any of this works.
What it does not do
- It does not prevent anything. It detects and it documents. Whether behaviour changes depends on what people do with the record.
- It does not remove the person. Events still go to someone who decides what to do. The software narrows what they have to look at; it does not make the judgement.
- It does not identify individuals. Detecting a person and recognising which person are different technologies with different regulatory obligations.
- It does not fix bad placement. The most common cause of a disappointing deployment is a camera pointed at the wrong thing, and that failure is invisible until you need the footage.
If you are evaluating platforms, our comparison of the main options covers where each fits, and what a system actually costs covers the part most vendors leave until the call.
Common questions
- What is AI video analytics in simple terms?
- It is software that looks at a video feed and works out what is in it and whether that matters. Instead of recording everything and hoping someone reviews it, it identifies specific things — a person in a restricted area after hours, a vehicle stopping where vehicles do not stop — and sends that event to a person with a clip attached.
- What is the difference between AI video analytics and motion detection?
- Motion detection triggers on any change in pixels, so rain, headlights, shadows and moving foliage all set it off. AI video analytics identifies what the moving thing is before deciding. That difference is why motion-based systems get muted after a fortnight of false alerts and analytics-based ones do not.
- Is intelligent video analytics the same as AI video analytics?
- Yes. The terms are used interchangeably, along with AI CCTV in British and Australian usage. There is no meaningful technical distinction between them — they all describe software that interprets video automatically rather than just recording it.
- Does AI video analytics use facial recognition?
- Not inherently, and ours does not. Detecting that a person is present is a different capability from identifying which person it is. Facial recognition is separately regulated in several US states and carries obligations that general detection does not. Ask any vendor to be explicit about which one they are selling you.
- Can it work with my existing cameras?
- Usually. Analytics is software and can run against cameras already installed. The limit is placement rather than compatibility: if the existing camera is not pointed at the thing you care about, or is too far from it, no software recovers that. New units mostly earn their place where there is no power or network.
- How accurate is AI video analytics?
- Detecting people and vehicles at reasonable range is highly reliable. Accuracy falls as targets get smaller in frame, which makes camera placement, framing and lighting far more predictive of results than which vendor's model is running. Ask for the imaging distance behind any accuracy claim.
Working out whether this fits your site?
Tell us what you need to know about and where the cameras are. We will tell you what is realistically detectable and what is not.