AI CCTV: What It Can Do With Cameras You Own

AI CCTV is software that watches camera footage and labels what is in it: a person, a vehicle, a fall, a forklift too close to someone. Alerts then fire on those events instead of on every change in the picture, and you can search hours of video in seconds. Most business CCTV can get it without new cameras, but how far each camera can "see" for the AI matters more than which AI you buy.

This guide covers what AI does with CCTV, how to tell if your cameras are good enough, the ways to add it, what it costs and where it fails, with links to a deeper guide where we have one. We build custom AI for CCTV, and we still say where a cheap boxed option is enough.

What AI does with CCTV footage

Every AI CCTV feature starts with object detection: a neural network (software trained on thousands of labelled images) finds people, vehicles and other objects in each frame. Rules and further models sit on top. The features below go from things every recent camera does well to things that need a model trained for your site.

The first two rows solve the problem most businesses actually have: CCTV that records everything and alerts on so much that nobody looks. If that is you, start there. We walk through each step up in more depth in the four levels of AI in security cameras.

Is your CCTV ready for AI? Four checks

You can answer these from the recorder's menu and a walk round the site, before anyone quotes you.

1. IP or analogue? IP cameras plug into a network switch with an Ethernet cable and can feed almost any AI. Analogue cameras (coax cable into a DVR, often labelled HD-TVI) cannot send a stream to other software. A DVR is the recorder for analogue cameras; an NVR is the IP version. You can swap the DVR for an AI one: Hikvision's AcuSense DVRs tell people and vehicles apart on analogue channels, though they run perimeter rules such as line crossing on only a few channels, so check the datasheet. Otherwise add an encoder (a box that turns an analogue feed into an IP stream) or replace the cameras.
2. Does it publish a standard stream? Almost every IP camera from the last decade offers RTSP (the protocol recorders use to pull video) and ONVIF (the standard for finding and controlling cameras). That is what lets an AI box read the same video as your recorder, whatever the brand.
3. Is each camera close enough to what you want caught? This is the check most buyers skip, and the next section works it out.
4. Can it see at night? A camera's infrared lights reach a set distance, 30 metres on the Hikvision turret we use as the example below. Beyond that the AI sees black unless the site is lit, whatever the brochure says about the model.

How far a camera can see, for a person and for the AI

CCTV designers use four levels of detail from the international standard IEC 62676-4, measured in pixels per metre at the target. Axis's white paper on the standard gives them: 25 pixels per metre to detect that someone is there, 63 to observe clothing, 125 to recognise someone you have seen before and 250 to identify a stranger. The further the person, the fewer pixels they cover.

We worked out the distance for each level on a very common UK business camera, the 4MP Hikvision turret DS-2CD2346G2H-IU with a 2.8mm lens, which Hikvision lists at 2688 × 1520 pixels and 100.2° wide. Then we did the same for its 640-pixel sub-stream, the low-resolution copy cameras send for phone viewing. Hikvision's own figures on that page are longer (detect 63 m, identify 6 m). Datasheet figures are usually worked out for the centre of the picture, where a wide lens packs its pixels most densely, as Axis's paper explains; ours spread the pixels evenly across the view, so treat them as the typical case and Hikvision's as the best case.

Two things follow. First, a wide camera covering a yard can tell you someone is there at 45 metres but cannot show who they are beyond about 4.5 metres. If you need faces or plates, you need a camera aimed at a choke point such as a gate or doorway, not better AI.

Second, check which stream the AI reads. Analysing a small stream saves a lot of computing, so many AI recorders and boxes do it. Frigate, a popular open-source AI recorder, tells users in its camera setup guide to detect on a low-resolution stream sized to the objects they want to catch. On a 640-pixel stream, the same camera reaches the "detect" level at about 11 metres instead of 45. The IEC levels are written for a person watching a screen, and models differ, but at 45 metres a person on that stream is about a dozen pixels tall, which is hard for any model. When AI "misses" someone at the far fence, this is usually why.

Four ways to add AI to CCTV you already own

Our AI security camera buying guide compares the first three routes in detail, with a 16-camera worked example. The software route ranges from free to bespoke. Frigate is open source under the MIT licence and runs on your own PC with an AI accelerator (a small add-on chip that runs the detection) or a graphics card; it suits a technical owner with a handful of cameras. A custom model, the route our computer vision team builds, is for the events no product detects.

Cloud platforms move recording and AI off site. Before you go that way, read how much upload 16 cameras need in our cloud CCTV costs and bandwidth guide. Verkada is the best-known cloud system, and its licences add up: our Verkada five-year cost breakdown works through a 16-camera site.

What AI CCTV costs

Hardware prices are UK retail including VAT, checked in November 2026; dollar figures use £1 = $1.34. Installation, cabling and drives are extra.

Our opinion: if person and vehicle alerts fix your problem, a few hundred pounds of AI recorder is the best money you will spend on CCTV. Pay for custom work only when the thing you need caught is specific to your business, or the estate is too big or mixed for boxed products.

How accurate is AI CCTV?

Vendors often quote one headline figure, such as "90% fewer false alarms". Treat any single number with care, because accuracy belongs to a camera on a site, not to a product. The same model that is near perfect on a lit car park struggles on a dark yard with a hedge in the wind.

The common failures, and what fixes them:

• Insects and spider webs at night. Infrared lights attract them, and a bug on the glass looks bigger than a person. Turn off the built-in infrared, light the area with a separate lamp mounted away from the camera, and keep the lens cover clean.
• Rain, snow and headlights. Better models filter most of it; a hood and a slight downward tilt help more.
• People too far away. See the distances above. Add a camera, do not tune the model.
• Behaviour changes after a firmware update. Note the version that works and re-test after every update.

The honest way to judge accuracy is to measure it on your worst camera. Run the AI for two weeks, log every alert as real or false, and check the recordings for events it missed. Our weapon detection for Angel Protection reached 97.3% detection accuracy because it was fine-tuned on images taken from CCTV angles for that one job; it now runs on more than 1,000 cameras.

AI CCTV, remote monitoring and police response

Many businesses pay a remote monitoring centre to watch alerts overnight. AI matters here because a centre's operators can only handle so many false alerts, and filtering them is what keeps the monthly fee sane.

If you want the police to respond to a camera alert in England, Wales or Northern Ireland, the system has to qualify for a police unique reference number. For detector-activated CCTV that means certification to BS 8418:2021, installed and maintained by a firm approved by a UKAS-accredited certification body such as NSI. A self-installed AI box that sends phone alerts is useful, but it is not that. Scotland runs its own policy.

Is AI CCTV legal in the UK?

Yes, under the same law as any CCTV: UK GDPR and the Data Protection Act 2018, as amended by the Data (Use and Access) Act 2025. You need a clear purpose, signs, a retention period and a way to answer requests for footage. The ICO says a data protection impact assessment (a written check of the risks and how you reduce them) is a legal requirement for surveillance likely to be high risk, which covers most video surveillance, including monitoring people at work.

Person and vehicle detection on cameras you already run rarely changes your legal position: same footage, sorted faster. Face recognition does, because identifying people from their faces is biometric processing, special category data under UK GDPR. So does using AI to judge how staff work. In the US the rules vary by state, with Illinois' biometric privacy law the strictest.

Where to start

1. Write down the three events you most want caught. "Someone in the yard after 10pm" is a product feature. "A forklift within two metres of a person" is a project.
2. Run the four checks above on the cameras that cover those events, and move or add cameras where the distance is wrong.
3. Try the cheapest route on your worst camera and measure it for two weeks.
4. Go custom only if the boxed route fails on something that matters. A computer vision feasibility study answers whether your cameras can do it before you pay for the build.

AI we have built for CCTV

For a UK CCTV security company we built real-time trespass and violence detection that runs on site: one RTX 5060 Ti handles 15 to 20 cameras and the footage never leaves the building.

For Angel Protection, a US school and hospital security firm, our weapon detection runs on more than 1,000 CCTV cameras, alerts about a second after a weapon appears and cut their hardware cost by 90%. Read the Angel Protection case study or how the system scaled to 1,000 cameras.

Our AI CCTV guides

• AI Security Cameras for Business: A Buying Guide: AI cameras, an AI recorder or an AI box, with prices and a 16-camera example.
• Cloud-Based CCTV: Real Costs, Upload Speeds and AI: the upload a cloud system needs, five-year costs and where the AI should run.
• Verkada Pricing 2026: Real 5-Year Cost in £ and $: cameras, licences and cheaper routes for a 16-camera site.

Questions

Is there such a thing as AI CCTV?

Yes. Most business cameras and recorders sold in recent years run person and vehicle detection, and separate AI software can add it to older IP cameras. "AI" on a box can mean anything from a motion filter to face recognition, so check what it actually detects.

How is AI used in CCTV?

Mainly to cut false alarms by alerting only on people and vehicles, to apply rules such as line crossing after hours, and to search recordings by description. Beyond that, trained models detect specific events such as falls, fights, weapons or missing PPE.

How much does AI CCTV cost?

From about £100 ($135) for one 4MP AI camera, or £177 to £390 ($240 to $525) for an AI recorder that adds person and vehicle alerts to existing Hikvision cameras. Verkada, the best-known cloud system, starts at $599 a camera before tax plus a $249 yearly licence. Custom detection starts with a £4,000 to £6,000 feasibility study.

Can I add AI to my existing CCTV?

Usually. IP cameras that publish an RTSP stream can feed an AI recorder or separate AI software without being replaced. Analogue cameras need an AI DVR, an encoder or new cameras.

Do AI cameras need Wi-Fi?

No. Business AI cameras are normally wired with PoE (one cable for power and data) and run detection on site. You need internet only for phone alerts, remote viewing or a cloud platform.

Is there free AI CCTV software?

Yes. Frigate is free, open source and widely used; it runs on your own hardware and works best with an AI accelerator or graphics card. You take on setup, updates and security yourself, which suits a technical owner more than a busy site manager.

Can AI CCTV recognise faces?

It can, but on most cameras only within a few metres: our worked example identifies people up to about 4.5 to 6 metres. In the UK face recognition processes biometric data, so you need a strong, documented reason and a DPIA before switching it on.