ANPR due diligence tests whether a parking operator's cameras and back office turn vehicle movements into correct parking charges, and whether the evidence behind each charge would survive an appeal, a DVLA audit and the 2026 Private Parking Code of Practice. In practice that means measuring capture and read rates per camera, testing how entry and exit reads are paired (the source of double-dip and phantom overstay errors), checking the age, ownership and connectivity of the camera estate, and confirming that every notice had a recorded manual quality check. There is no published UK read-rate standard for parking; the Home Office's police standard of 98% capture and 95% correct reads is the nearest reference point.
The wider context, from the KADOE rules to the parking technology stack, is in our parking sector guide. This article goes one level down, into the cameras and the evidence. It's the workstream I'd put most weight on when buying an ANPR operator, because read quality drives appeal rates, complaints, cancellations and regulatory exposure at the same time, and it's the one most often left to a vendor's summary slide.
A police camera only has to read a plate once. A parking charge needs two reads of the same plate, at the right times, matched to each other and to every payment made in between.
How ANPR parking enforcement works, end to end
A private ANPR car park turns a vehicle movement into a parking charge through a chain of steps, and an error at any one of them becomes a wrong charge.
- A camera at the entrance captures an image of the vehicle and its plate, with a time stamp.
- Software reads the plate into a vehicle registration mark (VRM), with a confidence score.
- A camera at the exit does the same.
- The back office pairs the entry and exit reads for the same VRM into a visit.
- It checks the visit against tariff payments (from machines, apps, phone lines and pre-books), permits, whitelists and any exemptions, and applies the consideration and grace periods.
- Visits that still look like a breach go to a person for a manual quality control check of the images.
- If the check passes, the operator requests the registered keeper's details from the DVLA through KADOE.
- A notice to keeper is posted. Where no notice was given to the driver on site, Schedule 4 of the Protection of Freedoms Act 2012 requires it to be delivered within 14 days beginning with the day after the parking ended, in England and Wales, if the keeper is to be held liable.
Then come appeals, reminders, debt recovery and occasionally court. Steps 1 to 5 are where the technology lives. Step 6 is where the 2026 code now draws a hard line, and step 7 is where a systemic error turns into a data protection problem, because a keeper request without reasonable cause is a misuse of DVLA data. Our article on KADOE covers that side.
What is a good ANPR read rate?
Two numbers matter, and vendors often blur them. Capture rate is the share of vehicles passing the camera that produce a read at all. Read rate (or read accuracy) is the share of captured plates read correctly.
The only published UK standard is for law enforcement. The Home Office's National ANPR Standards for Policing and Law Enforcement, version 3.5 dated July 2026, requires static systems to capture 98% of plates that meet the reflectivity standard and are visible to the human eye, and to read 95% of captured plates accurately. For multi-lane cameras reading vehicles travelling away from the camera, the capture requirement drops to 90%. The standard also sets out a test method: capture is assessed on at least 250 consecutive vehicles or two hours of traffic, read accuracy on at least 250 consecutive captured vehicles, in daylight and at night, with read compliance reviewed at least annually.
NASPLE says plainly that it doesn't cover ANPR used for purposes other than law enforcement, so no parking operator is bound by it. It is still the best reference point a deal team has, and the test method is worth copying.
Here is why the police figure is not enough for parking. A police camera only has to read a plate once. A parking charge needs two reads of the same plate, at the right times, matched to each other and to every payment made in between. If each camera independently achieves 98% capture and 95% accuracy, about 93% of passes produce a correct read, and only about 87% of visits get a correct read at both entry and exit (my arithmetic, assuming independent errors). Back-office matching rules, fuzzy matching and the manual check exist to close that gap. A target that relies on the camera alone is relying on luck.
Vendor claims need testing against that frame. One UK camera supplier's November 2025 newsletter markets an AI-based camera at "99%+ accuracy across UK plates", with no sample size, conditions or test method published. Which? said in an April 2026 piece on ANPR errors that the industry "boasts a 97% accuracy rate", again without a named source. Neither figure tells you anything about a particular retail park entrance at dusk in February. Ask for per-camera statistics measured on the target's own sites, and if they don't exist, that is the finding.
Double dips, phantom overstays and other ANPR enforcement errors
Most wrong ANPR charges come from a small number of failure patterns. Each leaves a signature in the data, which is why raw reads are worth more in diligence than any summary.
Double dip. A driver visits the same car park twice in a day. The cameras miss the first exit and the second entry, so the back office pairs the first entry with the final exit and records one long stay. The 2026 code names this directly: its guidance on the manual check says it is "particularly important for detecting issues such as 'double dipping'".
Phantom overstay from a missed exit. The exit camera misses a vehicle altogether, so the visit stays open until the vehicle's next entry, days or weeks later. On a site with 1,000 visits a day, a 2% missed-exit rate leaves 20 visits a day that look as if the vehicle never left.
Misread VRM. An O read as a 0, or a B as an 8, so the entry and exit don't match each other or don't match the payment the driver made. Under the code's keying-error clause, operators must have a documented procedure to avoid charging for accidental keying errors by drivers, including technology that matches entered VRMs against vehicles recorded on site.
Tailgating. Two vehicles pass the camera close together and the second plate is hidden. The code notes the manual check may reveal tailgating problems that suggest the camera should be moved.
Clock drift. Entry and exit cameras disagree on the time, or drift from the payment system, so a 2 hour 58 minute stay becomes 3 hours 4 minutes. The code requires equipment to be "synchronised", and the operator must keep a record of how it has guaranteed the accuracy of the time stamp where it relies on images.
Drive-throughs and passing traffic. Cameras that catch vehicles using the car park as a cut-through, or reading the road outside. Which? reported a driver who received four charges for a car park he had only driven past.
Cloned and altered plates. Not a camera fault, but it shows up in the same data, and the code's indicative data list already asks for appeals citing stolen or cloned vehicles.
Each of these is testable. In the data room, ask for open visits older than 24 hours by site, multiple same-day visits per VRM by site, charges cancelled on appeal by reason, and the distance in seconds between entry and exit camera clocks on a sample day.
The ANPR camera estate: lifecycle, ownership and connectivity
ANPR cameras sit outdoors on poles, on mobile data connections, often on finance and sometimes owned by somebody else. That makes the estate a balance sheet question as much as a technical one.
Start with ownership. For each site, establish who owns the cameras, signs, poles and connectivity, who pays for maintenance and replacement, and what happens to all of it when the landowner contract ends. Camera vendor finance agreements are a common place for change-of-control termination rights to hide.
Then age and support: build an asset register by model, installation date, firmware version and vendor support status, and check it against what is actually on the poles. A sample site visit is cheap. Cameras out of support can't be patched, and they are network-connected devices on public land.
Ofcom's advice for IoT device suppliers, updated in June 2026, says all four UK mobile network operators have completed their 3G switch-offs. 2G is next, on a timetable each operator sets: EE from May 2029, O2 from summer 2029 and the Vodafone 2G network during 2030, with 2033 the outer limit. Any camera, router or barrier controller still dependent on 2G or 3G is either already failing or has a known end date inside a typical hold period. Count them and cost the replacement.
The government's Private Parking Certification Scheme expects a surveillance technology policy covering how equipment is synchronised and maintained, and maintenance logs showing the date of each check, any issues and how they were resolved. If those logs don't exist today, they will have to by the first certification audit.
What does the 2026 code require of ANPR operators?
The Private Parking Code of Practice, laid before Parliament on 20 October 2026, turns most good ANPR practice into rules. Our article on the private parking code of practice 2026 covers the whole code. For the camera estate, five provisions matter.
| Code provision | What it requires | What to test in diligence |
|---|---|---|
| 7.1 Use of photographic evidence technology | Equipment fit for purpose, maintained to manufacturers' requirements and synchronised; design, security and maintenance information recorded for inspection | Maintenance logs, clock synchronisation method, asset register |
| 7.3(d) Manual quality control check | ANPR or CCTV images can only support a charge after a manual quality control check, including time stamp accuracy and the risk of keying errors | Check records per notice, time per check, who did it, rejection rate |
| Certification scheme, surveillance module | A record of the individual completing the quality check before a notice to keeper | Named user on every check in the audit trail |
| 6.3 Keying errors | A documented procedure, including technology matching entered VRMs against vehicles on site | Fuzzy-match rules, keying-error cancellations by site |
| 17.1 Operator data | Record and retain data for at least 36 months and supply it through the accredited association, from ANPR events and charges to appeal outcomes | Whether the back office can produce the indicative data list today |
The manual check is the one to dig into. "Manual" is easy to claim. Ask how many checks each person does per hour, what share of proposed charges they reject, and whether anyone is targeted or rewarded on volume, which the code separately bans for charges issued. A team approving 600 checks an hour is not checking. A rejection rate near zero suggests the same thing. A healthy process rejects a meaningful share of proposed charges, and the reasons for rejection are themselves the best map of where the camera estate needs work.
This is also the boundary for AI. Software can pre-screen every proposed charge, rank low-confidence reads and flag near-match payments, but the code requires a person to make the check. Our piece on AI in parking enforcement sets out where that line sits.
ANPR, GDPR and the ICO
The ICO's guidance on ANPR says a VRM is personal data in most circumstances. It expects a data protection impact assessment before deployment, showing the use is necessary and proportionate. It asks for databases and matching of sufficient quality to prevent mismatches, clear signage that ANPR recording is taking place, data sharing agreements with third parties, and retention for the minimum period necessary.
Its car park example is directly relevant. A gym enforcing a two-hour limit collects data on every vehicle, including those that leave in time; the ICO says keeping data on compliant vehicles longer than necessary is "unnecessary and excessive" and they should be deleted as soon as appropriate. Set that against the code's 36-month data requirement and there is an apparent conflict. The answer is aggregate counts kept long, personal data on compliant vehicles kept short, and evidence on charged vehicles kept for as long as the charge, appeal or claim is live. Ask the target to show you that its back office actually does this, not that its policy says so.
Automated decisions are the newer issue. Since February 2026, the Data (Use and Access) Act 2025 has replaced Article 22 of the UK GDPR with new Articles 22A to 22D. A decision is "based solely on automated processing" if there is no meaningful human involvement, and significant decisions taken that way need safeguards, including the ability to contest the decision and obtain human intervention. A parking charge issued with no real human check would sit uncomfortably with both the code and that regime. Any operator adding AI to the pipeline needs an updated DPIA.
For a deal, the data protection checks are: a DPIA per site type, a retention schedule and evidence that deletion runs, processing agreements with camera vendors, cloud hosts and debt recovery agents, signage that names the controller, and a log of who in the business can access images and keeper data.
What to request in the data room
| Request | Why |
|---|---|
| Per-camera capture and read statistics for the last 12 months, with method | The core quality measure; a network average hides the worst sites |
| Raw entry and exit reads for a sample of 10 to 20 sites over three months | Lets you rebuild visits and find double dips, open visits and clock drift yourself |
| The matching rules engine, its configuration by site and its change history | Shows how reads become charges, and when the rules changed |
| Manual check records: user, timestamp, outcome and reason, per notice | Code requirement, and the best evidence of a real check |
| Charges cancelled at first and second stage, by site and coded reason | Where evidence fails; free-text reasons are a finding in themselves |
| Asset register: model, install date, firmware, support status, connectivity, owner | Capex, 2G exposure and change-of-control risk |
| Maintenance logs and fault tickets, with time to repair | Certification evidence and real uptime |
| Camera finance and maintenance contracts | Change of control, termination and ownership on exit |
| DPIAs, retention schedule, deletion logs and processor agreements | ICO compliance |
| Last three DVLA audit reports and remediation evidence | ANPR controls are part of every audit |
| Clock synchronisation method and drift records | The time stamp is the evidence |
An ANPR test protocol for due diligence
Documents only take you so far. On an ANPR operator, I would run a short independent test alongside the technical due diligence, on a sample of sites chosen by the buyer rather than the seller: the busiest, the worst for appeals, a dual-lane entrance and one with poor lighting.
- Ground truth. At each sampled entrance and exit, record independent video for at least 250 consecutive vehicles, by day and at night, following the NASPLE method.
- Capture and read. Compare the system's reads with the video. Calculate capture rate and read accuracy per camera, separately for day and night.
- Pairing. For the same window, rebuild visits from raw reads and compare them with the back office's visits. Count open visits, double dips and mismatched pairs.
- Clocks. Compare entry, exit and payment system timestamps against a reference clock.
- Charge re-check. Take a random sample of around 100 recently issued charges from the sampled sites and re-check each one from the images, payment records and permits, without seeing the original decision. Count the charges you would not have issued.
- Manual check audit. For the same sample, pull the check record: who checked, when, how long it took and what they saw.
- Back-office replay. Re-run three months of raw events through the matching rules with the code's five-minute consideration and ten-minute grace periods, and measure how many charges disappear.
That takes days, not weeks, and turns a vendor's accuracy claim into a number for each site. Steps 5 and 7 are the ones that move price, because they say how much of today's charge income is either wrong or about to be ruled out. Treat any "AI" in the pipeline the same way, as a claim to test; our guide to AI washing in due diligence covers the general approach, and our technology due diligence checklist covers the rest of the back office.
If you are buying an ANPR operator or camera vendor and want the estate tested before the price is fixed, our technical due diligence team runs this protocol as part of the work, and the parking sector guide sets out how it fits the wider diligence.
Questions
How accurate are ANPR cameras in car parks?
There is no published UK accuracy standard for parking ANPR, and vendor figures are rarely backed by a test method. The nearest benchmark is the Home Office standard for police cameras, which requires static systems to capture 98% of plates visible to the human eye and read 95% of those correctly. Accuracy in a car park depends heavily on camera position, lighting, tailgating and plate condition, so it has to be measured per camera on site.
What is a double dip in ANPR parking?
A double dip happens when a vehicle visits a car park twice in a day and the cameras miss one exit and one entry in between. The back office then pairs the first entry with the last exit and records one long stay, which looks like an overstay. The 2026 Private Parking Code of Practice names double dipping as one of the errors the manual quality control check must catch before a charge is issued.
Is ANPR data personal data under UK GDPR?
The ICO says that in most circumstances a vehicle registration mark captured by a surveillance system is personal data, because it can be linked to an individual. Operators therefore need a lawful basis, a data protection impact assessment before deployment, clear signage, accurate matching, sharing agreements with processors and minimal retention. The ICO's own car park example says keeping data on compliant vehicles for longer than needed is unnecessary and excessive.
How long can a car park keep ANPR data?
There is no fixed period. The ICO expects data on vehicles of no interest, such as those that left within the time limit, to be deleted as soon as appropriate. Data supporting a charge can be kept while the charge, appeal or any court claim is live. The 2026 code separately requires operators to keep operational data, such as charges and appeal outcomes, for at least 36 months, which can mostly be met with aggregated records.
Can AI replace the manual check of ANPR images?
Not under the 2026 code. Images generated by ANPR or CCTV can only support a parking charge after a manual quality control check, including the time stamp and the risk of keying errors, and the certification scheme expects a record of the individual who completed it. AI can rank, pre-check and flag cases so the person checks faster and better, but a person must still make the check.