What Happens When an ANPR Camera Misreads a Plate
95% is the read accuracy figure in the Home Office National ANPR Standards for Policing and Law Enforcement, version 3.5, published in July 2026. A system built to that standard is still expected to get 1 in 20 of the plates it captures wrong, and the gap is deliberate. No camera reads every plate correctly, in every condition, on every pass, and the standard is honest enough to say so. For a car park the more useful question is what happens next: who notices, what the record shows, and how a fair result is reached for the vehicle sitting behind that read.
13 August 2026 / 7 min read
By Tim Marting, Head of International Business Development
How accurate is ANPR in practice?
The national standard for police ANPR sets 2 separate measurements. Capture is the share of plates a camera picks up at all. Read accuracy is the share of those captured plates it gets right. A static system must capture 98% of plates that meet the reflective requirements and accurately read 95% of what it captures.
Those 2 figures answer different questions, and a supplier quoting only one of them has told you half the story. A camera can read almost everything it sees and still miss a share of vehicles, because it never framed them in the first place.
The same standard sets a lower capture figure of 85% for a camera doing 2 jobs at once, general CCTV as well as plate reading. That is a design trade off rather than a fault. A camera covering a wide area cannot frame every plate as tightly as one pointed down a single lane, which is why dedicated ANPR cameras are specified separately where read volume matters.
Be precise about what this standard is, though. It governs policing and law enforcement systems, and no equivalent figure is imposed on a private car park by law. It is simply the clearest published benchmark available, and a fair yardstick to hold any system against.
The 3 reasons a plate read goes wrong
3 things send a read wrong. The plate itself may fall short of the legal standard, or be dirty, damaged or blocked. The conditions at the moment of capture, meaning angle, weather and light, can defeat a well placed camera. And a driver can type their own registration in incorrectly at a payment point.
Every plate fitted after 1 September 2021 has to meet British Standard BS AU 145e: a single shade of black, no background pattern, reflective material, black characters on a white front plate and black on a yellow rear one. The characters are specified to the millimetre, 79mm tall and 50mm wide apart from the number 1 and the letter I, with a 14mm stroke and 11mm between them.
Those rules exist because a plate that is hard for a person to read is hard for a camera to read. A vehicle with an incorrectly displayed plate can attract a fine of up to £1,000 and will fail its MOT.
Conditions matter just as much. The angle a vehicle takes through a lane, road spray in heavy rain, low winter sun straight into the lens, or a bike rack sitting across the plate can all defeat a camera for that one vehicle on that one pass, with nothing wrong with either the plate or the camera.
Then there is the part a person does by hand. At payment kiosks and in the Parka app a driver enters their own registration, and a dropped character happens the way it does anywhere someone types a reference number.
This is the part that shapes everything after it. A misread plate and a registration typed in wrong look identical once they land in the data. Both leave the same thing behind: a visit with no payment matched to it.
The plate that reads perfectly and is still wrong
A cloned plate defeats read accuracy completely, because nothing about the read is wrong. The characters are correct and the vehicle is not. GOV.UK describes cloning as copying the identity of a similar vehicle already on the road that has not itself been stolen. No read rate catches that on its own.
A cloned plate is not dirty, damaged or badly lit. It is clean, standard and entirely readable, which is rather the point of it. The safeguards that deal with a clone are therefore not optical at all. They sit in the quality of the record and the checks built around it.
The Information Commissioner's Office puts this directly in its ANPR guidance: both the cameras and any algorithms used to determine a match must be of sufficient quality to prevent any misidentification of a registration mark, and the databases behind them must be kept up to date, accurate and of sufficient quality to prevent mismatches. The ICO also treats a registration mark as personal data in most circumstances, with the pre deployment privacy groundwork covered step by step in our guide to ANPR cameras.
That is why Article 5(1)(d) of the UK GDPR reaches the record of every read. Personal data must be "accurate and, where necessary, kept up to date; every reasonable step must be taken to ensure that personal data that are inaccurate, having regard to the purposes for which they are processed, are erased or rectified without delay". The duty is identical in Ireland under the equivalent wording in the EU GDPR, with the Data Protection Commission as regulator.
The practical defence is the picture. A read that stores an image of the whole vehicle, and not only the characters, gives a site something to compare when a plate and a car do not belong to each other.
What has to happen before a charge is issued
Under the sector's joint industry code, a photograph cannot be the basis of a parking charge unless at least one image carries a clear record of the vehicle's registration mark, and the image has passed a manual quality control check that includes the accuracy of its timestamp. A person has to look at it.
The Code is version 1.1, effective 17 February 2025 with its foreword updated on 13 April 2026, and it is the single rulebook shared by the sector's 2 accreditation bodies. That manual check is the step that catches most bad reads before they reach anybody at all.
Images used as evidence cannot be altered, other than to blur a face or another vehicle's registration mark. Cameras have to be operated overtly, and in a reasonable, consistent and transparent manner, which in practice begins with clear signage telling a driver that a site uses camera technology before they park. Operators must also follow a documented procedure to avoid issuing or enforcing a charge over an accidental keying error.
There is a real difference in how sites fail here. At a site with barriers and access control a failed read announces itself at the lane, so it gets solved over the intercom while the driver is still sitting there. At a barrier free site nothing visible happens at all, and the first anyone knows about it is a notice arriving days later.
Neither is wrong. They fail differently, and knowing which failure a site is signing up for is part of choosing the system. Parka supplies and supports both, and we would rather help a site pick its failure mode on purpose than discover it later.
See what happens to a read that fails
Book a short walkthrough of the exception path, from the moment a plate does not read to the moment it is resolved.
Book a demo →The audit trail, and what the appeals numbers show
POPLA, the independent appeals service for British Parking Association members, received 104,349 appeals in the year to 30 September 2025, passing 100,000 for the first time. It decided 67,680 of them, allowing 14,578 and refusing 53,102. A further 39,522 were recorded as not contested.
POPLA defines a not contested appeal as one where either party withdrew before the appeal concluded. Operators withdrew in 34,502 of those cases and appellants in 5,020. The report does not record a reason case by case, so the plain reading is the useful one: more than a third of the appeals raised were settled before an adjudicator had to decide anything.
That is what a good record at the front end is for. Parking charge records have to be kept for 36 months. A charge can be appealed within 28 days, and the operator has 28 days to respond or to acknowledge that a decision is still open. Once an appeal reaches the appeals service, enforcement has to stop.
The Appeals Charter is specific about typing mistakes. A minor keying error, 0 instead of O, I instead of L, 1 instead of I, should see the charge cancelled where the tariff was paid. Where several characters are wrong, POPLA's published expectation is a reduced administration charge, usually up to £20. POPLA also expects an operator to compare the registration a driver entered against similar looking plates its own cameras recorded.
Since February 2025 the Charter has gone further at camera monitored sites: a driver who pays for the full duration of their stay before leaving should not receive a parking charge. Every one of those steps rests on a clean, timestamped record of the visit, which is what sits inside the operator platform.
What to ask a supplier about accuracy
The question that separates suppliers is not the headline accuracy figure. It is how a site handles the reads that fall outside it, and how often the rate is measured after installation. A number quoted once at the point of sale describes a camera. A number remeasured on a schedule describes a site.
The standard is specific about how a read rate should be established in the first place. An assessment needs a minimum of 250 consecutive vehicles observed over at least 2 hours. Read rates are then reviewed annually, and every 6 months for cameras that do not keep supporting imagery, using samples of 50 vehicles or more. A rate measured on installation day and never looked at again is not worth much a year later.
So the questions are practical ones. How is the read rate measured, and how often? When a plate does not read, who sees it, and how quickly? Is there a written keying error procedure that matches the Code? Does the stored image show the whole vehicle rather than the characters alone? And what can a driver actually do about it in the moment?
Parka works with operators in whatever way suits the site, whether that is cameras, barriers, kiosks or all of them together. We would rather agree how exceptions get handled while a car park is still on the drawing board than after the first notice goes out.
Read accuracy will never reach 100%, and it does not need to. A site that catches its rare bad reads early, and settles them without a fuss, earns something no specification sheet can show: drivers who come back without giving the car park a second thought.
Common questions
Is ANPR 100% accurate?
No system is, and the national standard says as much. It asks police ANPR systems to accurately read 95% of the plates they capture, alongside a separate capture figure that is itself below 100%. The gap is built into how the technology is measured, which is why a well run site plans for occasional wrong reads rather than assuming they will not happen.
What happens if a driver types their registration in wrong?
It depends on the size of the mistake. A minor slip, such as 0 instead of O or I instead of L, is one the operator is expected to identify and cancel where the tariff was paid. Where several characters are wrong, POPLA's published expectation is a reduced administration charge, usually up to £20, rather than the full parking charge.
Does a misread automatically cancel a parking charge?
Not automatically. The Code requires a manual quality control check of the image before a charge is issued at all, which is where most bad reads should be caught. Beyond that, an appeal is assessed against the evidence, including the operator's own images and payment records, and enforcement is suspended while the appeals service considers it.
How often should an operator check their camera read rate?
The national standard sets a clear rhythm. Compliance is documented when the system is first installed, using a minimum of 250 consecutive vehicles observed over at least 2 hours. After that the read rate is reviewed annually, or every 6 months for cameras that do not keep supporting imagery, using samples of 50 vehicles or more.
More insights
Your sites already have the answers
A 30 minute demo shows the occupancy, payment and enforcement picture your car parks generate every day.