All guides

Facial Recognition CCTV in the UK: Law, Accuracy, Cost

By Iu Ayala 13 min read Updated

Facial recognition CCTV compares each face a camera sees with a watchlist and alerts a person when one matches. In the UK it is legal for police and for businesses, but a business must show it is necessary, write a risk assessment first and put up signs. In an independent test of the software the Met uses, the default setting missed 11 to 15 in 100 of the people it was looking for and wrongly flagged about 1 in every 6,000 to 7,000 passers-by.

This guide is for a business deciding whether to add face matching to its cameras. The last section covers your rights if a camera has scanned your face. We do not sell a facial recognition product or run a watchlist. We build computer vision on existing CCTV, which can include face matching, and where a view is ours we say so. We are engineers, not solicitors.

Jump to: how it works, who uses it, the law, what a business must do, accuracy, the camera it needs, cost, does it cut crime, alternatives, your rights, questions.

How does facial recognition CCTV work?

Every system, from a police van to a shop door, does the same five things:

  1. Find a face. The software spots a face in the picture. This is face detection, and by itself it identifies nobody.
  2. Measure it. The face is turned into a string of numbers called a biometric template.
  3. Compare it. The template is scored against every face on a watchlist: people the operator wants to know about.
  4. Apply a threshold. If a score is above a set level, the system raises an alert. Otherwise nothing happens, and well-run systems delete the template at once.
  5. A person decides. Someone looks at both pictures and chooses whether to act. The software suggests a match. It does not confirm one.

The same technology is used in four ways, and the law treats them differently:

KindWhat it doesTypical user
Live (LFR)Checks everyone who passes a camera against a watchlist, as they passPolice vans and fixed cameras, shop entrances
RetrospectiveSearches a still from recorded footage against a database, after an incidentEvery UK police force
Operator-initiatedAn officer photographs one person and checks who they areSouth Wales and Gwent Police; a Met trial
One to oneChecks that you are who you say you are: a door, a phone, a passport gateAccess control, where the person takes part knowingly

The rest of this guide is mostly about the first kind, because it scans people who have done nothing and did not choose to take part.

Who uses facial recognition in the UK?

Police. A briefing for Parliament counted 13 of the 43 forces in England and Wales using live facial recognition in March 2026, with a national rollout and 40 more vans planned. The Met scanned 3,147,436 faces in the year to September 2025 and made 962 arrests from the alerts. Fixed cameras piloted in Croydon led to 173 arrests between October 2025 and March 2026, according to the Home Office. Forces are expected to tell the public where and when they use it, so your force's website is the place to check for locations.

Shops. The best-known supplier of facial recognition cameras to UK shops is Facewatch, which shares a watchlist of suspected offenders between subscribing retailers. Sainsbury's had it in 55 stores in July 2026 and said it would reach more than 200 by Christmas, The Grocer reported, naming Home Bargains, Southern Co-op, Budgens and Costcutter as other users.

Employers and venues. Some have used it to clock staff in or let members through a door. The regulator has acted most firmly here: it ordered Serco Leisure to stop in 2024.

Border and immigration officers. Immigration Enforcement, part of the Home Office, uses live facial recognition at UK ports.

Yes, with conditions. No UK law bans it, and none was written for it. Businesses are governed by UK GDPR and the Data Protection Act 2018, enforced by the Information Commissioner's Office (ICO). Police rely on common law powers, the same data protection rules, human rights law and their own published policies.

A face template used to identify someone is biometric data, which UK GDPR puts in its most protected class, special category data. Ordinary CCTV needs a lawful reason. Face matching needs that plus a second, narrower condition, and it has to be necessary: if a less intrusive method would do the job, the law expects you to use that.

Four decisions show where the line sits:

  • Bridges v South Wales Police, 2020. The Court of Appeal ruled the force's use unlawful: officers had too much discretion over who went on the watchlist and where cameras went, and the force had not checked the software for bias. For you: write down who can be listed and where you scan, and test for bias.
  • The ICO and Facewatch, 2023. The ICO closed its investigation after Facewatch narrowed its watchlist to repeat or serious offenders, and said the decision was "not a blanket approval". For you: catching thieves can be a lawful purpose, and each deployment is judged on its own facts.
  • The ICO and Serco Leisure, 2024. Ordered to stop scanning the faces and fingerprints of more than 2,000 staff to record attendance at 38 leisure centres, because ID cards or fobs would do. For you: convenience is not necessity, and staff cannot freely say no to an employer.
  • Thompson v Met Police, April 2026. The High Court upheld the Met's live facial recognition policy, which limits cameras to crime hotspots, protective security operations and places where specific intelligence says a wanted person is likely to be. For you: a written policy with real limits is what makes use defensible.

The rules are about to change. On 4 November 2026 the Home Office announced a specific law for police facial recognition, an independent watchdog that could stop a force using it, and "a separate legal code for the private sector" covering businesses that use it to tackle crime, "including shops". No text has been published yet. The ICO has also marked its facial recognition guidance as under review because of the Data (Use and Access) Act 2025. Anything you sign today should allow for both.

In the US there is no federal law on face matching, and the rules vary by state. In December 2023 the Federal Trade Commission banned the pharmacy chain Rite Aid from using facial recognition for five years. It said the company had not tested the system's accuracy, had used poor-quality pictures and had not trained the staff who acted on its alerts, which wrongly flagged shoppers as thieves.

What must a business do before using facial recognition?

The ICO's guidance for small businesses on crime answers "can I use facial recognition?" with a yes and a warning: its lawful use "has a high bar because it scans everyone in the environment, not just those already suspected". Together with the ICO's facial recognition checklist, it asks for the following, in writing, before the system is switched on:

  • Proof that it is necessary. What problem you have, what you tried first and why less intrusive options would not work.
  • A data protection impact assessment (DPIA). A written risk assessment, reviewed before every new deployment.
  • A lawful basis and a special category condition. For crime prevention the condition is usually "substantial public interest", which needs a short document called an appropriate policy document.
  • Watchlist rules. Each person is added case by case, for a reason you can evidence, and removed when the reason ends.
  • A human check that counts. Someone trained reviews every match before anyone is approached. A system that acts with no human input falls under the stricter rules of UK GDPR Article 22 on automated decisions.
  • Measured error rates. You can tell true matches from false ones, record both and change the settings if either is too high.
  • A camera that is good enough. The checklist asks for an "appropriate resolution" and "sufficient lighting", tested.
  • Retention limits. Delete the template of anyone who does not match at once, and set a written period for watchlist entries and alerts.
  • An equality check. The checklist asks you to consider an equality impact assessment under the Equality Act 2010, because error rates can differ between groups.
  • Clear signs saying facial recognition is in use, and a way for people to ask what you hold on them.

Police forces and councils must also follow the Surveillance Camera Code of Practice. A private business is not bound by it, though it is a sensible standard to work to.

Consent rarely works as the legal basis. The ICO's own example is a gym that makes face entry a condition of membership: that consent is not freely given. Offer a card as an alternative and it can be. Our guide to CCTV and GDPR covers the duties every camera system has before face matching is added, and CCTV in the workplace covers monitoring staff.

How accurate is facial recognition CCTV?

It is accurate enough to be useful and wrong often enough to hurt people. The answer is two numbers that pull against each other: how many of the people on the list it catches, and how many people who are not on the list it flags. One setting, the match threshold, trades one for the other.

The best public data is the National Physical Laboratory's 2023 test of the software the Met and South Wales Police use. At the default threshold of 0.6 it caught 85% of the people it was looking for and wrongly flagged about 1 in 7,000 of everyone else, against a watchlist of 10,000 faces. The report's own summary, which uses a looser face detection setting, gives 89% and 1 in 6,000.

Raising the match threshold means fewer wrong alerts and more missed people. At 0.56: 94 of 100 people on the watchlist raise an alert and 65 of the 100,000 people who are not on it are wrongly flagged. At 0.58: 88 of 100 people on the watchlist raise an alert and 27 of the 100,000 people who are not on it are wrongly flagged. At 0.60: 85 of 100 people on the watchlist raise an alert and 14 of the 100,000 people who are not on it are wrongly flagged. At 0.62: 82 of 100 people on the watchlist raise an alert and fewer than 4 of the 100,000 people who are not on it are wrongly flagged. At 0.64: 79 of 100 people on the watchlist raise an alert and none of the 100,000 people who are not on it are wrongly flagged. No setting catches everyone and flags no one.

100 people on the watchlist walk in

100,000 people who are not on it walk in

Everyone else walks in without an alert.

0.56 0.58 0.60 default0.62 0.64
← looseststrictest →
listed people caught
85%
listed people missed
15%
strangers flagged per 100,000
14

At a door with 1,000 visitors a day, wrong alerts: about 1 a week

  • On the watchlist, alert raised
  • On the watchlist, missed
  • Not on it, wrongly flagged
Numbers from the National Physical Laboratory's 2023 test of NEC NeoFace V4, the live facial recognition software used by the Metropolitan Police and South Wales Police, with wrong alerts scaled to a watchlist of 10,000 faces (Table 5, face detection settings a). At 0.62 the report gives an upper bound; at 0.64 it saw no wrong alerts. The headline figures the report quotes for the default, 89% and 1 in 6,000, use a looser face detection setting. Another product, camera or watchlist will give different numbers.

At the strictest setting tested, nobody was wrongly flagged and 21 in 100 listed people walked past unnoticed. At the loosest, 94 in 100 were caught and about 1 in 1,500 passers-by was flagged. There is no setting that catches everyone and flags no one. A vendor that quotes a single accuracy percentage has left one of the two numbers out.

To put the rates in a shop: 100,000 faces is 100 days of trade at 1,000 visitors a day. At the default setting the wrong alerts come to about 1 a week, each one a customer who did nothing.

Three things move these numbers in real use:

  • Watchlist size. Wrong alerts rise with the number of faces on the list. The report puts them at 1 in 6,000 with 10,000 faces on the list and 1 in 60,000 with 1,000. A shop's own list of 40 people is safer than a shared national one.
  • Picture quality. The Met's review of its own wrong alerts blamed poor lighting, poor angles and faces partly covered by clothing or other people. One was an identical twin.
  • The person at the end. Software only suggests. In the reported cases of shoppers wrongly thrown out of UK shops, Facewatch blamed human error in the store and not its software: staff had added the wrong person to a list or approached the wrong person on the shop floor.

In the street, the Met reported 10 false alerts from 3,147,436 faces in the year to September 2025, about 1 in 315,000. Read the other way, 0.5% of its 2,077 alerts were wrong, and eight of the ten people wrongly flagged were Black, which the Met's annual report calls a very small sample. The report also says the Met runs at 0.64, the strictest setting in the figure above. The lab saw no wrong alerts there. The street produced ten.

Is it biased?

It depends on the software and the setting, which is why you should ask for test results by skin tone, sex and age. The 2023 test above found no statistically significant difference between ethnic groups or sexes at the default threshold or stricter, and more wrong alerts for Black subjects at looser settings. It also caught under-20s less often than adults. A 2025 test of a different product, the search tool on the Police National Database, found that at one setting it wrongly matched 0.04% of white subjects, 4.0% of Asian subjects and 5.5% of Black subjects. In its November 2026 announcement the Home Office said a replacement it has tested "will be available for police forces to roll out from next year".

Can my existing CCTV cameras do facial recognition?

Usually not from where they are. A camera placed to watch a room sees faces too small and from too steep an angle. Face matching needs a camera placed for faces.

Motorola's Avigilon publishes its guidelines for a face match, which are typical: a well-lit scene, people facing the camera, the camera tilted no more than 30 degrees from horizontal for best results, and a minimum number of pixels across the face. The last column is our own sum for a common 4MP wide-angle camera (2.8mm lens, a view about 100° wide).

Match confidencePixels across the facePixels per metreFurthest a face can be from the lens of a 4MP wide-angle camera
High804572.5 m
Medium502863.9 m
Low402294.9 m

Now add the angle. A camera 2.8 m up looks down at a face 1.6 m up, and the angle only drops to 30 degrees once the person is 2.1 m away. Any closer and the camera looks down too steeply. Yet 2.1 m from the wall is also where the face drops below 80 pixels. So for a face at that height, a ceiling camera with a wide lens gets a high-confidence match in a strip of floor less than 10 cm deep, and a medium one out to 3.7 m. Height changes it sharply: a face 1.8 m up gets a strip about 50 cm deep, and a face 1.5 m up gets none. Police systems are stricter still: in the 2023 test the London cameras ignored any face with fewer than 60 pixels between the eyes.

Side view of a camera 2.8 metres up a wall. Closer than 2.1 metres along the floor it looks down on a face more steeply than 30 degrees. Beyond 2.1 metres the face has too few pixels for a high-confidence match, and beyond 3.7 metres too few for a medium one. That leaves a strip less than 10 centimetres deep for a high-confidence match on a face 1.6 metres up.

  • Too close: steeper than the maker advises
  • High-confidence match
  • Medium confidence only
  • Low confidence or none
Our own geometry, drawn to scale: a 4MP camera with a 2.8mm lens mounted 2.8 m up, a face 1.6 m from the floor, and Avigilon's published guidelines of 30 degrees of tilt and 80 pixels across the face for a high-confidence match, 50 for a medium one. Distances are along the floor from the wall.

In practice that means one extra camera per entrance, at about head height, with a narrower lens, facing people as they walk in, with light on their faces and none behind them. Ask the supplier how the camera copes with a bright doorway behind the person (wide dynamic range), how it performs at night, and for the measured error rates of the matching software. The recorder and the other cameras can stay as they are. Our guide to CCTV for retail shops covers where the ordinary cameras go and how close each needs to be.

How much does facial recognition CCTV cost?

Few suppliers publish a price, so treat every figure here as a starting point for a quote. There are three ways to buy it:

RouteWhat you pay forPublic price
A subscription service with a shared watchlistA camera at the door, a monthly fee per store, a contract termFacewatch advertised £199 a month in 2020 and we could not find a current published price
Face matching inside a camera or video system you runCapable cameras or a server, a software licence, an installer's timeQuote only: the UK trade sellers we checked hide camera prices behind a login
Software built for your siteA feasibility study, then a build on your own camerasOur prices: £4,000 to £6,000 for the study, £15,000 to £20,000 for a working first version

With the second and third routes you hold the biometric data, so every legal duty above is yours. With the first, the provider usually takes that role for the matching, but your staff still write the reports and make the stops.

The costs that are not on the quote are the DPIA, which may need a solicitor, the training, and the time of whoever answers each alert. Budget for a wrong stop too. In the reported UK cases retailers apologised with vouchers of £25 to £100, and The Grocer reports one legal challenge.

Does facial recognition reduce crime?

The published evidence comes from the organisations using it. Sainsbury's says 92% of the offenders it identified and banned did not come back, and described a "seismic" fall in theft at its trial stores. The Met reported, through the Home Office, that overall crime in the part of Croydon with fixed cameras fell 10.5% on the year before. Neither figure comes with a method, a comparison area or an independent check, and we could not find an independent evaluation of facial recognition in UK shops.

An alert tells you a listed person walked in. It does not tell you a theft was prevented. If you trial it, compare stock loss and recorded incidents with the same weeks a year earlier and with a similar site that does not have it, and count the wrong alerts as carefully as the right ones.

What can you use instead of facial recognition?

The law makes you answer this in writing before you scan a face, so it is worth answering properly. Each option does a different job:

OptionWhat it tells youWhat it cannot tell youLegal bar
Facial recognitionA known person has walked inAnything about a stranger, or what anyone is doingHighest: biometric data from everyone
Behaviour detection on your CCTVSomething is happening: a fight, a person in a closed areaWho the person isLower: nobody is identified, a DPIA is still needed
Staff alerts with a still imageWho to look out for, shared with colleagues or a local schemeAnything automaticMiddle: pictures of suspected offenders have their own data rules
Cards, fobs or PINsWho came through a doorWhether the card is with its ownerLowest

We build behaviour detection, so allow for that. For a UK CCTV security company we built real-time trespass and violence detection that runs on a computer on site, so footage stays in the building, and one mid-range graphics card handles 15 to 20 cameras. It works on what people do. There is no watchlist to get wrong and no biometric data to protect. It cannot tell you that the person who just walked in stole from you last week. Our guide to CCTV video analytics lists what the software can and cannot detect, and the same on-site design runs weapon detection on more than 1,000 cameras for a US security company.

Our view: facial recognition earns its legal cost when a few known people cause most of the harm and you have trained staff to answer each alert. If you cannot name the people you would put on the list, it will not find them for you. For the other layers a shop needs, see what a retail security system is made of.

What are your rights if a camera scans your face?

In a shop or other business. You cannot opt out at the door: the camera scans everyone who enters, and the only way to avoid it is to stay out. The business must tell you with a sign. You can ask what it holds on you with a subject access request (a written request for your data, answered within one month), ask for wrong data to be corrected or deleted, and complain to the business and then to the ICO. If you are stopped, ask who operates the system: the watchlist is often held by a provider, and the shop holds the report that put you on it.

In the street. Police forces are expected to tell the public when live facial recognition is running, unless there is a critical and time-sensitive threat. If there is no match, forces say the face data is deleted automatically. When there is an alert, an officer near the camera looks at the two pictures and decides whether to speak to you.

Questions

Which UK supermarkets use facial recognition?

Sainsbury's has published the clearest numbers: 55 stores in July 2026 and more than 200 planned by the end of the year. The Grocer also names Home Bargains, Southern Co-op, Budgens and Costcutter. A shop that uses it must show a sign at the entrance.

What is the difference between face detection and facial recognition?

Face detection finds that a face is in the picture. Facial recognition works out whose face it is by comparing it with stored faces. Detection alone does not create biometric data about an identified person, and many cameras use it only to focus or to save a clear snapshot.

What confuses facial recognition?

Poor light, steep camera angles, low resolution and anything covering the face. Those were the causes the Met gave for its false alerts in 2025, along with one identical twin. Old reference photos also lower the match score.

Can I use facial recognition on my home CCTV?

Some home cameras have it: Google Nest calls it familiar face detection and you switch it on yourself. If your camera sees only your own property, data protection law does not apply. If it sees the street or a neighbour's garden, the ICO says it does, and identifying passers-by by face would be very hard to justify.

Can I use facial recognition for door entry or clocking in?

Only if people have a real choice. Checking one face against one enrolled picture at a door is less intrusive than scanning a crowd, but it is still biometric data. Offer a card, fob or PIN as an equal alternative, and do not make the face scan a condition of the job or the membership. Serco Leisure was ordered to stop because staff had no such choice.

Does the UK government use facial recognition?

Yes. Police forces use it live and on recorded images, and Immigration Enforcement uses it at ports. The Home Office funds the police vans: £11 million to triple their number, it said in November 2026.

Do I need permission to install facial recognition cameras at my business?

There is no licence to apply for. You need the written justification, the DPIA and the signs described above, and you must pay the ICO data protection fee like any CCTV operator. If your DPIA finds a high risk you cannot reduce, you must consult the ICO before you start.