AI Licence Plate Recognition for Low-Emission Zones

Why low-emission zone enforcement is harder than it looks

A low-emission zone (LEZ) is a simple promise: cleaner air where people live, work and breathe. It is easy to announce and genuinely hard to enforce. The moment a city draws a boundary, it inherits a round-the-clock operational problem — thousands of vehicles crossing in and out, at every hour, each one either allowed or not. Get enforcement wrong and the zone fails twice: compliant drivers feel unfairly punished, and non-compliant ones learn the rules are not real.

Manual checks do not scale, and the cheap automated alternatives are brittle. A human officer cannot watch a junction at 3 a.m. in driving rain. Off-the-shelf cameras can — until conditions turn. Night, low winter sun, rain on the lens, spray from a lorry, a plate half-hidden behind a tow bar, the smear of motion blur as a car accelerates away: each quietly drops recognition accuracy, and every missed or misread plate is lost revenue, a wrongly issued fine, or a complaint that lands on someone's desk. This is a computer-vision problem in the real world, not in a lab, and the difference is everything.

This page is an honest capability deep-dive, not a victory lap. We describe how we would build automatic number-plate recognition (ANPR) for an LEZ, the technology behind each step, and what realistic accuracy looks like across the industry. We built this as a proof of concept for the Municipality of Antwerpen (District Antwerpen). The figures further down are sector benchmarks, not our own delivered results — we report your real numbers from a pilot on your own footage.

How the system works, step by step

Reliable enforcement is not one clever model; it is a short pipeline where each stage does one job well and hands clean output to the next. The goal is a single defensible decision per vehicle, with the evidence behind it.

• Detect. A fine-tuned YOLOv8 detector locates vehicles and licence plates in the frame, even through partial occlusion, motion blur and variable light. Training on hard, region-specific examples — wet roads, night scenes, awkward angles — separates a model that works in summer from one that works in a Dutch November.
• Read (OCR). Before any character is read, the plate crop is cleaned up: contrast normalisation, binarisation and skew correction straighten and sharpen the image. OCR then reads the characters, constrained by plate-format and regional rules so impossible combinations are rejected rather than guessed.
• Cross-reference. The recognised plate is validated against vehicle-registration data to confirm make, year, emission category and ownership — the step that turns a string of characters into a policy decision: is this vehicle allowed in this zone, today?
• Decide and log. A transparent rule engine evaluates eligibility against the zone's current rules, flags violations, and — crucially — logs every decision with an auditable record and role-based access. Operators get dashboards suited to their role; auditors get a trail explaining exactly why each call was made.

The principle throughout is fairness by traceability: a fine that cannot be explained should never be issued, so every output carries its evidence and reasoning.

The technology and approach behind the accuracy

The headline accuracy of any ANPR system is decided long before deployment, in the data and the engineering choices. We treat detection and recognition as separable problems so each improves independently, and tune the pipeline to the conditions the camera will actually face.

On the detection side, a modern single-stage detector such as YOLOv8 is fast enough for real-time use and accurate enough to find small, angled or partly hidden plates — provided it is trained on the right material. Generic models trained on clean daytime imagery fall apart on the edge cases that dominate real roadsides. So the work is less about model architecture and more about disciplined machine learning: curating hard examples, augmenting for rain, glare and blur, and validating against the site's messy reality, not a tidy benchmark set.

On the reading side, OCR quality depends heavily on preprocessing. Correcting skew, normalising contrast and applying format constraints removes whole classes of error — an O misread as a 0, a B as an 8 — because the system knows what a valid plate can look like. The cross-reference step adds a second, independent check: a plate that reads cleanly but is not in the register is a signal to flag, not a fine to issue.

Two engineering decisions shape everything else. The first is edge versus cloud: running inference at or near the camera keeps latency low and reduces how much imagery has to travel — which matters for both performance and privacy. The second is integration: structured results and the audit log are exposed via API, the heart of any serious AI implementation, so the system drops into the workflow a city already runs rather than demanding a new one. Behind both sits solid data engineering and, where a custom operator console is needed, focused application development.

Who it is for, and the business value

This capability is built for the organisations responsible for clean-air and access policy: municipalities and public-sector bodies running low-emission or zero-emission zones, but also operators of congestion schemes, restricted areas, ports and depots where knowing which vehicle entered, and whether it was allowed, carries real financial weight.

The return on investment is rarely abstract. Reliable, around-the-clock enforcement does three measurable things at once:

• It protects revenue and fairness. Every plate missed or misread is a fine that should have been issued but was not, or one that should not have been issued but was. Tightening accuracy improves both the books and public trust.
• It reduces manual cost. Officers stop watching footage and chasing ambiguous cases, and spend their time on the genuinely difficult decisions. The system handles the volume; people handle the judgement.
• It creates defensible decisions. When every flagged violation arrives with the image, the read, the cross-reference and the rule that triggered it, appeals resolve faster and are far less likely to succeed against you. Auditability is not overhead — it is the asset.

Crucially, the value is measurable in your own terms. We do not ask you to take a generic accuracy figure on faith; we instrument the pilot so you see read-accuracy, capture rate and false-positive rate on your own traffic, against your own baseline. That makes the ROI conversation concrete rather than hopeful, and you can see indicative scopes on our pricing page.

An honest word on the numbers

It would be easy to print a single impressive accuracy figure and let you assume it is ours. We will not. The benchmarks below come from ANPR industry reporting and studies, not from a Crux Digits deployment. Across the sector, real-world plate-read accuracy typically lands between 90% and 98%, with capture rates around 98% and read rates near 95%; under controlled conditions, above 99% is achievable. Those are useful planning anchors — and nothing more.

Real-world performance is shaped by your site: camera placement and resolution, lighting, traffic speed, the mix of plate formats, and how aggressively the zone must catch edge cases. That is exactly why we report verified read-accuracy from your own roadside footage, not from a lab or a brochure. The capability is proven; the precise number is yours to measure, and ours to report honestly.

How we would run a pilot for you

We start small and prove value before anyone commits to a full rollout. A focused pilot turns the open question — "how well would this work here?" — into a measured answer in weeks, on your own data.

• Scope and baseline. We agree the zone, the rules, the success metrics and the current manual baseline, so there is a clear yardstick to beat.
• Collect and tune. We gather representative roadside footage — including the hard conditions — and tune detection and OCR to your plates, angles and lighting.
• Measure honestly. We report read-accuracy, capture rate and false-positive rate on your footage, and show the audit trail behind sample decisions, so you see the reasoning, not just the score.
• Decide and integrate. A clear go or no-go follows. On a go, structured results and the auditable log connect via API into the enforcement systems your operators already use.

Throughout, two non-negotiables hold. Plate data is personal data, so we design with data minimisation, retention limits and access control from the start; the legal basis and retention policy are set by the operating authority. And the system stays accountable: every decision is logged, explainable and reviewable. If you run an LEZ or any access-controlled zone and want enforcement that is consistent, fair and fully auditable, book a free consultation — or see how the same computer-vision discipline plays out on a production line in our concrete crack detection case study.