Inside an AI number-plate recognition pipeline

Automatic number-plate recognition (ANPR) looks simple: point a camera at a road and read the plates. In practice it is a pipeline of models and rules, and every stage matters.

The stages

1. **Vehicle detection:** a model finds vehicles in each frame.
2. **Plate localisation:** a second step finds the plate on the vehicle.
3. **Character recognition:** a deep-learning model reads the characters.
4. **Post-processing:** results are checked against valid plate formats, and readings across several frames are combined into one confident result.

Real-world challenges

• Glare, rain, night-time and headlights.
• Dirty, damaged or non-standard plates.
• Fast vehicles and motion blur.
• Different plate formats, including regional and diplomatic plates.

Engineering for accuracy

Camera choice and placement matter as much as the model: the right lens, shutter speed and infrared illumination give the model a readable image. Train and test on images from the actual sites, and measure accuracy by condition — day, night, rain.

From reads to records

Each read should carry a timestamp, location, confidence score and image, so it can be verified. Low-confidence reads can be routed for human review.

Privacy and retention

Plate data is personal data. Limit who can query it and how long it is kept.

See our AI transport systems.