CCTV Analytics: What Works, What Fails and Why
CCTV analytics is software that watches camera footage for you and raises an alert, a count or a search result when it sees something you defined: a person in the yard after hours, a van parked too long, a queue building. Spotting people and vehicles is dependable today. Behaviour, crowds and faces are much less so.
This guide covers what CCTV video analytics can detect, how reliable each type is, what it needs from the camera, and how to test it before you pay. We build custom analytics for CCTV, but most of what follows applies just as well to the free analytics already inside your cameras.
What is video analytics in CCTV?
Video analytics (also sold as video content analysis, VCA, or intelligent video analytics, IVA) is any software that turns video into events. The label covers three generations of technology that behave very differently, and a spec sheet rarely says which one you are getting.
| Generation | How it decides | What that means on site |
|---|---|---|
| Motion detection | Enough pixels changed between frames | Fires on rain, headlights, shadows and spiders. Fine for deciding when to record, poor for alerts outdoors |
| Rule-based analytics | Follows moving blobs and checks their size, speed and direction against a line or zone | Line crossing and intrusion on cameras from the 2010s. Better than motion, but a swaying branch of the right size still counts |
| Deep learning (AI) | A neural network, software trained on labelled images, recognises what the object is | Rules fire only for a person or a vehicle. This is the step that removes most false alarms |
Brand names map onto these. Hikvision's "Smart Event" and Dahua's "IVS" are rule-based line and zone detection; AcuSense (Hikvision) and WizSense (Dahua) add the deep learning filter for people and vehicles. If your recorder has line crossing but still alerts on foxes and headlights, you have the second generation, not the third.
How CCTV analytics works
Every system, from a £100 camera to a server room, runs the same four steps:
- Read a stream. The software takes video from the camera, often a low-resolution copy and only a few frames a second. Frigate, a popular open-source AI recorder, analyses 5 frames a second by default and says in its camera setup guide that this "is correct for almost all cameras".
- Detect objects. A model marks each person, vehicle or other object it was trained on in each frame.
- Track them. The software links detections between frames, so it knows this is the same person who was by the gate two seconds ago.
- Apply a rule. You draw a line or zone and set hours. Crossing it, staying in it too long or being counted through it becomes an alert, a number on a dashboard or a bookmark in the recording.
Most failures come from step one or step four: a camera that gives the software too few pixels, or a rule drawn across a public pavement. The model is rarely the weak point for people and vehicles.
The generations also work together. For Angel Protection, a US school and hospital security firm, we built weapon detection where cheap processors run a plain motion check on every stream and only frames with movement reach the graphics card that runs the AI. That design cut their hardware cost by 90% and now runs on more than 1,000 cameras. The weapon detection case study has the detail.
Types of CCTV analytics, and how far to trust each
Vendor pages list these as equals. They are not. The last column is our judgement, from the detection systems we have built and from what the manufacturers' own manuals admit.
| Analytic | What it does | How far to trust it |
|---|---|---|
| Intrusion and line crossing | A person or vehicle enters a zone or crosses a line in set hours | High in a quiet area with a well-placed camera. Buy this first |
| Camera tamper | The view is blocked, moved, defocused or lost | High. Free on most cameras; switch it on |
| Number plates (ANPR, or ALPR in the US) | Reads plates and checks them against a list | High on a dedicated camera aimed at a gate; poor on a wide security view |
| Search | Finds "white van" or "person in red" across recordings | Good. A wrong result costs seconds, so it does not need to be perfect |
| Loitering and parked too long | The same object stays in a zone past a time limit | Good in quiet scenes. In busy ones the tracker swaps people and the timer resets |
| People counting | Counts people through a doorway | Good with a camera looking down at the door; poor from a general security camera |
| Occupancy, queues, crowd density | Estimates how many people are in an area | Fair. Use it for trends and thresholds, not exact counts |
| Abandoned or removed object | Something was left behind or taken away | Fair in an empty, steady scene; unreliable where people stand about |
| Face recognition | Matches faces against a watchlist | Good only close up and face-on, and it is biometric data in law |
| Behaviour | Fights, falls, concealing stock, climbing | Low out of the box. Needs a model trained and tested for the event |
| Safety rules (PPE, forklift near a person) | Checks a site rule is being followed | Low to fair. Usually a custom project |
| Tracking across cameras | Follows one person from camera to camera | Low as automation. Useful as a search aid that suggests look-alikes |
There is outside evidence for that order. The UK Home Office used to benchmark analytics with a footage library called i-LIDS. Its named event detection scenarios were single-camera rules from this table: sterile zone (an area nobody should be in), parked vehicle, abandoned baggage and doorway surveillance. Its tracking test across multiple cameras was dropped, in the Home Office's words, "due to a lack of submissions", and it called the scenario "very challenging". The Home Office stopped distributing the library in 2014, so for an ordinary business the practical benchmark is now a trial on your own site.
Manufacturers are candid in their manuals too. Axis lists hard-hat detection in its Object Analytics as a beta feature, next to line crossing and time-in-area rules that it treats as finished.
Behaviour is where custom work comes in. For a UK CCTV security company we built real-time trespass and violence detection. We benchmarked two kinds of video model for it and chose the smaller one, X-CLIP, which needs about 1GB of graphics memory and 0.02 seconds per check, so one RTX 5060 Ti graphics card covers 15 to 20 cameras.
What analytics needs from the camera
The manuals give hard numbers that the brochures leave out. Axis publishes the clearest set in its Object Analytics user manual, and they are a fair guide for other brands:
- Size. A person must be at least 4% of the picture height to be detected at all, and Axis recommends 8% to avoid misses. For a vehicle it is 3% and 6%.
- Time. An object has to move in view for at least 2 seconds. Someone who darts across a narrow gap between two buildings may never be seen.
- Light. If the scene is lit by lamps and not by the camera's own infrared, Axis recommends at least 50 lux across the whole detection area.
- Angle. Not looking straight down, picture level with the horizon, and people roughly upright. Legs and wheels should be visible, so a hedge that hides people below the waist hurts more than you would expect.
The size rule turns into distances. We worked them out for a typical UK business camera, the 4MP Hikvision turret DS-2CD2346G2H-IU, from the fields of view on Hikvision's product page, for a person 1.7 metres tall standing square to the camera on flat ground.
| Target | 2.8mm lens | 4mm lens |
|---|---|---|
| Person at 8% of picture height (dependable) | up to 21 m | up to 26 m |
| Person at 4% (the limit, in perfect conditions) | up to 41 m | up to 52 m |
| UK number plate at 130 pixels wide | up to 4.5 m | up to 6.3 m |
Two lessons. First, the rule is a share of the picture, so more megapixels do not push it out; a narrower lens does. If the fence is 40 metres from the camera, a wide 2.8mm lens puts it at the edge of what the analytics can do, and no setting will fix that.
Second, plates need their own camera. Axis's licence plate reader manual asks for a single-row plate to be at least 130 pixels wide. A UK plate is about 52 centimetres across, so a wide security camera only manages that within about 4.5 metres. Our worked distances for detecting and identifying people cover the same camera for human viewers.
Axis's Object Analytics manual also lists what triggers false alarms:
- A partly hidden vehicle. A small van appearing from behind a wall is tall and narrow, so it can be classed as a person.
- Insects and spiders on the lens, drawn by the camera's own infrared lights at night.
- Car headlights combined with heavy rain.
- Human-sized animals, and strong light throwing shadows.
Why "98% accurate" tells you very little
Some installer pages quote accuracy of 98% or more without saying what was counted. The trouble is that almost nothing happens on a security camera, so a system can be right nearly all the time and still bury you in false alerts. Here is a made-up but realistic fortnight on one yard camera, watched 12 hours a night at 5 frames a second.
| System alerted | System stayed quiet | |
|---|---|---|
| Someone was there | 18 real alerts | 2 missed events |
| Nobody was there | 70 false alerts | 3,024,000 frames analysed in all, nearly all of them correctly ignored |
- Frames judged correctly
- 99.9%
- Alerts that were false
- 80%
- Real events caught
- 90%
A vendor can truthfully call this system 99.9% accurate. The person receiving the alerts sees 5 false ones a night and learns to swipe them away, and 2 of the 20 real visits were never flagged. Multiply by 16 cameras and nobody reads the alerts by the second week.
Ask a supplier for these three numbers instead, measured on a site like yours:
- False alerts per camera per night. Our opinion: more than one or two after tuning and people stop looking.
- Missed events. Out of every real event in the test, how many produced no alert. This needs someone to check the recordings, which is why it is rarely quoted.
- The conditions. Day or night, what distance, what weather. A figure from a floodlit forecourt says nothing about a dark yard with a hedge in the wind.
When we report 97.3% detection accuracy for the Angel Protection weapon model, it is for one object, on a model fine-tuned with images from CCTV angles. A number without that context is marketing.
False alerts cost more than attention if a monitoring centre (an alarm receiving centre, or ARC) or the police are involved. The first objective of the police chiefs' security systems requirements in England, Wales and Northern Ireland is "to reduce the number of false calls passed to the police", and every false alert takes up a monitoring centre operator's time. If you use an ARC, ask which analytics it accepts alerts from before you buy: the alert has to arrive with a video clip the operator can verify.
Where analytics runs: camera, recorder, server or cloud
| Where | Good at | Limits |
|---|---|---|
| In the camera (edge) | Person and vehicle rules with no extra box and no bandwidth | Small chip, so a fixed menu of analytics; set up one camera at a time |
| In the recorder (AI NVR) | Adds the people and vehicle filter to older cameras of the same brand | Often only on some of its camera inputs (channels); check the datasheet |
| On a server on site | Any camera brand, bigger models, custom detection, footage stays on site | A computer with a graphics card to buy and look after |
| In the cloud | Many sites managed from one screen, nothing to maintain | Upload speed (13 to 48 Mbps for 16 cameras, depending on the provider), a fee per camera, and video leaving the building |
You do not need the cloud for analytics. Most business systems run them on site and use the internet only to send the alert. Our comparison of AI cameras, AI recorders and AI boxes prices the first three for a 16-camera site, and the cloud CCTV guide works out the upload a cloud system needs.
What CCTV analytics costs
CCTV analytics costs less than most quotes suggest, because the dependable features come bundled with the hardware. Prices are UK retail including VAT, checked in November 2026, with dollars at £1 = $1.34.
- Built into the camera: free. A 4MP Hikvision AcuSense turret with person and vehicle rules is £100 to £135 ($135 to $180). Axis says its Object Analytics "comes preinstalled on compatible Axis network cameras at no extra cost".
- AI recorder: £177 to £390 ($240 to $525) for 8 to 16 cameras.
- Cloud platforms: a licence per camera per year. Verkada's is $249 (about £186) on top of the camera.
- Custom analytics: our feasibility study is £4,000 to £6,000, credited to a £15,000 to £20,000 build that takes 4 to 6 weeks.
The AI CCTV cost table has the full list. Our advice is to use up the free analytics before paying for anything: many sites own AcuSense or WizSense cameras with the rules never switched on.
Which analytics to switch on first, by type of site
| Site | Start with | Then consider |
|---|---|---|
| Yard, depot or construction site | Person and vehicle intrusion inside the fence, out of hours | ANPR at the gate; tamper alerts on every camera |
| Warehouse | Intrusion on dock doors and the perimeter at night | Counting at dock doors; custom rules for forklifts near people |
| Shop | Search, so finding an incident takes minutes | Footfall and queue counts from a camera above the door |
| Office or school | Out-of-hours intrusion at entrances and car parks | Loitering near entrances; occupancy for room use |
| Vacant property | Intrusion with alerts to a monitoring centre | Tamper alerts, because cameras there get covered or turned |
The same cameras can also count and measure for the business, not only guard it. Factories check products on the line, which we cover in our piece on computer vision for quality control. For Dendri Design, a logistics firm, our machine vision in the warehouse raised throughput by 40%. Pointing analytics at staff changes the legal picture, which the law section below covers.
How to choose CCTV analytics and a supplier
Work down this list and stop at the first row that passes your test.
| Option | Examples | Choose it when |
|---|---|---|
| Analytics in the cameras or recorder you own | Hikvision AcuSense, Dahua WizSense, Axis Object Analytics | You need person and vehicle rules and the cameras are recent |
| Analytics software on your own server | Frigate (free, open source); analytics add-ons for video management software such as Milestone or Genetec | Mixed camera brands, or you want search and rules in one place |
| Cloud platform | Verkada, or a cloud service added to existing cameras | Many small sites, good upload and nobody to look after a server |
| Custom model | Built for your event on your footage | No product detects what you need, or off-the-shelf analytics failed the test |
Whoever you buy from, ask these before you sign:
- Can we trial it on our worst camera for two weeks before we commit?
- How many false alerts per camera per night should we expect, and how many real events will it miss?
- Which stream does it analyse, at what resolution and how many frames a second?
- Does the analysis run on site or in the cloud, and where is footage stored?
- Is there a licence per camera or per analytic, and what happens if we stop paying?
- Who retunes the zones when the site changes, and what does that cost?
CCTV analytics and UK law
Analytics on CCTV falls under UK GDPR and the Data Protection Act 2018, like the cameras themselves. Detecting that a person crossed a line at 2am rarely changes your position: same footage, sorted faster. The duties that already apply to your CCTV still apply, and three uses of analytics add to them.
- The basics. A written purpose for each camera, a lawful basis (usually legitimate interests for security), clear signs saying who is recording and why, a retention period you can justify, and a way to answer requests for footage.
- A data protection impact assessment (DPIA). This is a written check of the risks and how you reduce them. The ICO's video surveillance guidance makes it a legal requirement where the processing is likely to be high risk, which includes monitoring people at work. Update it when you switch on a new kind of analytic.
- Face recognition. Identifying people from their faces is biometric processing, special category data under UK GDPR. You need an extra legal condition under Article 9 on top of the lawful basis, and a DPIA before you start.
- Analytics aimed at staff. Counting how long someone spends at a workstation is worker monitoring. Tell staff what is measured and why, and expect to justify why a less intrusive method would not do.
- Decisions about people. An alert is a prompt for a person to look, not proof. Do not ban a customer or discipline an employee on an alert nobody reviewed.
Our summary of UK CCTV law covers the cameras themselves. US readers: there is no single federal rule, and Illinois is the strictest state on biometrics.
How to test analytics on your own site
You can do this with the analytics you already own, before anyone quotes you.
- Pick one camera and one event. The worst camera, and the event you most want caught, such as a person inside the fence between 10pm and 6am.
- Draw the zone tight. Keep the pavement, the road and the tree out of it. Most false alerts come from a zone that includes somewhere people are allowed to be.
- Walk it at night. Have someone in dark clothes walk the far edge of the zone, then crouch, then move along the fence. Note what was caught and at what distance.
- Log two weeks. Mark every alert real or false, and write down the cause of each false one. A few causes, often one tree, one light or one stretch of pavement, explain most of them.
- Fix the view before the software. Move the camera, narrow the lens, add a light, trim the hedge. Then decide whether you need better analytics at all.
If the event you care about is not on any product's menu, or off-the-shelf analytics fail this test on something that matters, that is when a custom system earns its cost. A feasibility study on your own footage answers whether it can work before you commit to a build. Our AI CCTV guide covers the routes to add it to cameras you own.
Questions
What is AI video analytics software?
AI video analytics software analyses video automatically and turns it into alerts, counts or searchable records. In CCTV it runs inside the camera, in the recorder, on a server or in the cloud, and ranges from simple motion detection to AI that recognises people, vehicles and specific events.
What is the difference between motion detection and video analytics?
Motion detection reacts to any change in the picture, including rain and shadows. Video analytics recognises what moved and applies a rule to it, so you can alert only on a person crossing a line after hours. Motion detection is still useful for deciding when to record. Our four levels of AI in security cameras walks up from one to the other.
How does video AI work?
A neural network trained on labelled images marks the objects in each frame. Software then tracks each object between frames and checks it against rules you set, such as a zone, a direction or a time limit. A match becomes an alert or a count.
Can CCTV analytics work with my existing cameras?
Usually. Many cameras sold since about 2020 already have person and vehicle analytics that were never switched on. Older network (IP) cameras can feed an AI recorder or separate software. Analogue cameras need an AI recorder built for them, or an encoder, a small box that converts their signal.
How accurate is CCTV analytics?
It depends on the camera more than the product. Person and vehicle detection is dependable when the person fills about 8% of the picture height and the scene is lit. Ask for false alerts per camera per night and missed events, not a single percentage.
How much does CCTV analytics cost?
Basic person and vehicle analytics come free inside cameras from about £100 ($135) and AI recorders from £177 ($240). Cloud platforms charge per camera per year, $249 (about £186) for Verkada. Custom analytics start with a £4,000 to £6,000 feasibility study.
Does video analytics replace CCTV operators?
No. It decides which few minutes a person should look at. Someone still has to judge the alert and act, and any decision that affects a person, such as accusing them of theft, should be made by a human who has watched the footage.
Does CCTV analytics need the cloud?
No. Analytics in the camera, the recorder or a server on site work without sending video anywhere. You need an internet connection only for phone alerts, remote viewing or a cloud-managed platform.