Car parks, company sites, hotels and municipalities need to know which vehicle came in, when it left and whether it is allowed. Automatic number plate recognition (ANPR/LPR) reads the plate from a camera image and replaces tickets, barriers, cash and manual patrols with a plate-based session. Done right it cuts queues and operating cost; in the EU the plate is treated as personal data, so purpose limitation, signage and short retention are part of the design, not an afterthought.
Problem & value
- Today: paper tickets, RFID cards, barrier-and-cash machines, or patrol staff who write down plates. Lost tickets, fraud, queues at peak, and manual permit checks are the pain points.
- AI value: plate-based ticketless entry/exit, pay-by-plate and pay-by-app, automatic whitelist access for staff, residents and delivery fleets, overstay detection, occupancy counts, and reconciliation of payments against sessions for enforcement.
- Why AI: modern detectors plus OCR read plates in rain, at night and at angles where classic rule-based OCR failed; multi-country plate formats and non-standard fonts are handled by trained models rather than hand-written rules.
- Benefit figures: vendors report high read rates on well-installed lanes, but results depend on camera position, lighting and plate quality. Measure on your own lanes before committing to a "touchless" target [1][2].
Variants
- Gated car park / garage: entry and exit lanes, barrier control, payment on exit.
- Ticketless open car park or street parking: cameras at entry and exit points, pay-by-plate via machine or app.
- Corporate / residential access: whitelist of employee, tenant and visitor plates, visitor pre-registration.
- Parking enforcement (public): handheld or vehicle-mounted cameras check permits and paid sessions; a human officer issues the notice. Legal basis is national and differs for public authorities and private operators.
- Fleet, logistics yards, petrol stations, EV charging: gate check-in, dwell time, fuel-pay matching, charger reservation.
- Private use: home gate or driveway opener with a local camera and allow-list, processed locally.
- Out of scope here: road-network flow analytics (VIS-012), speed/offence enforcement (VIS-054).
How it works
- camera (IR / global-shutter, trigger by loop or motion)
- plate detection (object detector)
- crop, rectify, enhance
- character recognition (OCR / sequence model) + country/format classification
- post-processing (format rules, multi-frame voting, confusable characters O/0, B/8)
- decision (whitelist / booking / tariff lookup)
- barrier, display, payment, alert
- retention and deletion job; audit log
- Detection and OCR: a YOLO-class detector locates the plate; a compact recognition network reads characters. Multi-frame voting across the vehicle track raises accuracy.
- Edge inference: the lane controller or camera does the work on-site (Jetson, Hailo, camera SoC); only plate text and event metadata go to the backend, which supports data minimisation.
- Human-in-the-loop: low-confidence reads go to a review queue or to a manual intercom/ticket fallback; enforcement notices are always checked and issued by a person; a documented appeal path exists.
- Make/model/colour matching can add a plausibility check (plate-swap or cloned plates), usually as an optional second signal.
Data preparation
ANPR needs camera images at lanes or gates plus the lists behind them: whitelists, permit holders, bookings, reservations and tariffs. These typically sit in the parking or access system, in Excel files at the facility manager, in the booking or hotel system, in the ERP for staff permits and in paper permit folders. Typical gaps are plates with typos, several plates per customer, outdated permits, duplicate records, missing validity dates and unclear consent for storing plates.
Preparation steps: clean the permit list into one table of plate, holder, validity and permission (PREP-011); normalise plate formats (spaces, umlauts, country codes); remove expired entries; align retention periods and deletion rules with the legal basis (PREP-025); pseudonymise exports used for analytics (PREP-022). Test with 200 to 500 real images in day, night and rain conditions (PREP-018) and check camera angle, shutter speed and lighting.
Rough share of effort: low, about 15 to 25 percent, mostly list clean-up and camera setup. Ongoing upkeep of permits and bookings matters more than model data.
Implementation paths
Which path for whom
- SMB: Path A, because mature ANPR parking systems with payment and barrier integration exist at SMB prices and the process is standard.
- Enterprise: Path B, because large operators must connect plate reads to their own booking, permit, tariff and payment systems across many sites.
- Developer: Path C, because open plate detectors and OCR run on HW1 edge devices and allow custom integrations.
- Public: Path C, because municipalities need data sovereignty, tight retention control and compliance with procurement rules; certified products (Path A) only where available.
Path A — Buy (recommended for most SMBs)
- Parking-management suites with LPR: Skidata, Scheidt & Bachmann, Designa and Parkl (EU-based vendors), Amano McGann, Flowbird, plus Kapsch TrafficCom (EU) for tolling and city use. Many operators also use pay-by-plate cloud services (e.g. Parkopedia-like and municipal apps).
- Camera and LPR specialists: Adaptive Recognition (Hungary, EU), Vaxtor (Spain, EU), Plate Recognizer (offers SDK and cloud), Genetec AutoVu, Hikvision and Axis ANPR cameras/analytics.
- Check: EU hosting or on-prem processing, support for your countries' plates, configurable retention, access logging, DPIA template, open API for existing barrier and payment systems.
- Choose A when: a standard gated or ticketless site, no custom workflow; vendors typically supply cameras, installation and service.
Path B — Build on cloud APIs
- Plate Recognizer Snapshot / Stream (cloud or on-prem Docker), Azure AI Vision / Google Vision OCR for a fallback read, or a frontier VLM (Claude, Gemini, GPT, Mistral) to resolve hard crops.
- Run the detector on-site; send only cropped, low-resolution plate crops to a cloud API in an EU region.
- Orchestration: MQTT or webhooks to barrier controllers, payment gateways (Stripe, Adyen, Mollie), and booking systems.
- Choose B when: you integrate with an existing property or booking platform and want fast iteration.
- Caveat: a VLM is slower and costlier per read than a dedicated model; use it as a fallback only.
Path C — Self-host (open source)
- Models/tools: Ultralytics YOLO (AGPL-3.0 or commercial licence, check terms), fast-alpr with open-image-models and fast-plate-ocr, PaddleOCR, EasyOCR; OpenALPR legacy code is no longer maintained.
- Runtime: ONNX Runtime or TensorRT on Jetson Orin / Hailo-8 / Coral; RTSP ingestion with GStreamer or Frigate/Viseron-style NVR for home use.
- Hardware: HW1 handles several lanes per edge box; HW2 if you also run a VLM fallback centrally.
- Ops effort: camera positioning and lighting dominate; plan on-site tuning per lane.
Path D — Custom training
Worth it for unusual plate formats (e.g. Gulf, motorcycle, temporary or historic plates), extreme camera angles, or very high volumes. Needs a few thousand labelled plate crops per format from your own lanes, including night, rain, dirt and glare. Reviewer corrections from the exception queue are the natural training data. Public datasets exist but rarely match your site conditions.
State of the art (2026)
- Two-stage pipelines (detect, then recognise) remain the production standard; end-to-end recognition transformers and compact OCR models run in real time on edge hardware.
- VLMs are used as a fallback for degraded crops and for make/model checks; research shows they can support vehicle surveillance tasks, but they are not yet a replacement for tuned LPR on latency and cost [3].
- Failure modes: dirt, glare, tilted or obscured plates, similar characters, non-standard fonts and frames, cloned plates, plates with different country conventions, motion blur at speed. Plate recognition tells you which plate was seen, not who the driver is: keeper lookup is a separate, legally regulated step.
- Accuracy claims from vendors are tied to ideal conditions; test with your own footage and report per-lane read rates and false accepts at the barrier.
- Trend: pay-by-plate, permit-less virtual parking, EV-charger integration and "free-flow" multi-storey cars parks.
Shared resources & combinations
- Same stack (object-detection + OCR + edge-runtime): VIS-018 industrial OCR, VIS-012 traffic analytics, VIS-008 people counting, VIS-010 site security share edge boxes, RTSP ingest and the NVR.
- Privacy layer: VIS-030 anonymisation for stored footage and for sharing clips.
- Adjacent but different legal regime: VIS-054 enforcement cameras; VIS-016 face-based access control (more sensitive; see Art. 5 notes in that file).
- One edge GPU box can serve plate reading and occupancy counting for the same car park.
Effort, cost & KPIs
- Path A: weeks; hardware, installation and per-lane or per-transaction fees dominate.
- Path B: 1–3 months; cloud calls are minor relative to camera hardware and installation.
- Path C: 2–4 months for a robust multi-lane solution; maintenance of lighting, cleaning and calibration.
- KPIs: read rate and correct-read rate per lane (day/night/weather), false-open rate at the barrier, manual interventions per 1,000 sessions, payment match rate, queue time, enforcement-notice appeal and cancellation rate, retention-deletion compliance (100% on schedule).
People & change
Involve management for the business case, the facility or parking manager as process owner, and IT for barrier, payment and booking integration. The DPO is needed early: plates are personal data and the legal basis, signage and retention must be settled before launch. Legal or compliance reviews contract terms, parking fines and liability. Customers should be informed by clear signs and a privacy notice; suppliers (system integrator, payment provider) need an AVV.
Works council: depends. If staff car parks or employee permits are covered, or logs show arrival times, the system may be suited to monitor behaviour and §87(1) no. 6 BetrVG applies; otherwise customer-only parking does not touch it. Inform the council in good time under §90(1) no. 3 when work processes change. DPO: a DPIA is likely for large-scale or public-area use (GUIDE-004, GUIDE-005).
AI literacy: support the staff who handle exceptions with training on misreads and manual override (GUIDE-006). Pilot at one car park (GUIDE-009). Typical concerns are tracking of movements and wrong fines; address them with short retention, no profiling and a human check before any penalty is issued (GUIDE-008).
Risks, regulation & compliance
- EU AI Act: minimal risk for ordinary parking and access control. Note the Art. 5 ban on real-time remote biometric identification for law enforcement does not apply to plates (they are not biometric data), but combining plates with face recognition or tracking people across sites can shift the legal analysis. Public-authority use for law enforcement is separately regulated; check national police law. AI literacy (Art. 4) applies [4].
- GDPR: plates are personal data when the operator can reasonably link them to a keeper or driver (identifiability in the sense of CJEU Breyer, C-582/14). You need a legal basis (typically legitimate interest, Art. 6(1)(f), for private operators; legal duty or public task for authorities), purpose limitation, clear signage and information, data minimisation, short retention (delete non-matching reads quickly), access control and logging, processor contracts, and usually a DPIA [5][6].
- Retention: keep only what is needed for billing or enforcement; German supervisory authorities treat short deletion periods for video as the rule. Check your regulator's current guidance.
- Keeper data requests: requesting keeper details from vehicle registers is regulated nationally; claims for contractual penalties by private operators have been subject to court rulings in several Member States.
- Misuse: do not repurpose parking data for tracking or marketing; no cross-site watch-lists without legal basis.
- Security: camera networks and plate databases are attractive targets; segment networks, patch, encrypt.
- Mitigations: edge processing, hashed plate storage where feasible, strict role-based access, regular deletion audits, human review of every sanction.
Getting started
- Define purpose, legal basis and retention in a short DPIA before buying cameras.
- Survey the site: lane geometry, lighting, plate heights, vehicle speed, countries of origin.
- Pilot with one lane (Path A or fast-alpr on a Jetson) and log every read against ground truth for two weeks, incl. night and rain.
- Integrate with barrier, payment and whitelist; define the manual fallback (intercom, ticket).
- Set up signage, deletion jobs and access logs; review KPIs monthly.
Sources
- Adaptive Recognition — How to make sure your traffic solution is GDPR-ready — https://adaptiverecognition.com/blog/traffic-transportation/how-to-make-sure-that-your-traffic-solution-is-gdpr-ready/
- Plate Recognizer — GDPR for ANPR — https://platerecognizer.com/gdpr-compliance-for-anpr/
- Video-based Vehicle Surveillance in the Wild: License Plate, Make, and Model Recognition with Self Reflective Vision-Language Models (arXiv 2508.01387) — https://arxiv.org/pdf/2508.01387
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- EDPB — Guidelines 3/2019 on processing of personal data through video devices — https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-32019-processing-personal-data-through-video_en
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- fast-alpr (open-source ALPR pipeline) — https://github.com/ankandrew/fast-alpr
- PaddleOCR — https://github.com/PaddlePaddle/PaddleOCR
- Ultralytics YOLO — https://github.com/ultralytics/ultralytics