FLOCK automatic number plate recognition
page 11 / 11
=== Plate inconsistency and jurisdictional differences ===
Many ANPR systems claim accuracy when trained to match plates from a single jurisdiction or region, but can fail when trying to recognize plates from other jurisdictions due to variations in format, font, color, layout, and other plate features. Some jurisdictions (particularly in the US) offer vanity or affinity plates, which can create many variations within a single jurisdiction.
From time to time, US states will make significant changes in their license plate protocol that will affect OCR accuracy. They may add a character or add a new license plate design. ALPR systems must adapt to these changes quickly in order to be effective. Another challenge with ALPR systems is that some states have the same license plate protocol. For example, more than one state uses the standard three letters followed by four numbers. So each time the ALPR systems alarms, it is the user's responsibility to make sure that the plate which caused the alarm matches the state associated with the license plate listed on the in-car computer. For maximum effectiveness, an ANPR system should be able to recognize plates from any jurisdiction, and the jurisdiction to which they are associated, but these many variables make such tasks difficult.
Currently at least one US ANPR provider (PlateSmart) claims their system has been independently reviewed as able to accurately recognize the US state jurisdiction of license plates, and one European ANPR provider claims their system can differentiate all EU plate jurisdictions.
=== Accuracy and measurement of ANPR system performance ===
A few ANPR software vendors publish accuracy results based on image benchmarks. These results may vary depending on which images the vendor has chosen to include in their test. In 2017, Sighthound reported a 93.6% accuracy on a private image benchmark. In 2017, OpenALPR reported accuracy rates for their commercial software in the range of 95-98% on a public image benchmark. April 2018 research from Brazil's Federal University of Paraná and Federal University of Minas Gerais obtained a recognition rate of 93.0% for OpenALPR and 89.8% for Sighthound, running both on the SSIG dataset; and a rate of 93.5% for a system of their own design based on the YOLO object detector, also using the SSIG dataset. Testing a "more realistic scenario" involving both plate and reader moving, the researchers obtained rates of less than 70% for the two commercial systems and 78.3% for their own.
=== Limitations of legacy LPR systems ===
In some contexts, the term legacy LPR is used to describe older licence plate recognition systems that rely solely on black-and-white optical character recognition (OCR) of the plate itself, without capturing broader contextual data such as vehicle type, colour, or surrounding road conditions. These systems can be prone to higher false-positive rates, may struggle with new plate formats, and are generally limited in their ability to detect complex violations such as wrong-way parking, misuse of loading zones, or the presence (or absence) of parking permits visible only inside a vehicle's windscreen.
== See also ==
Anti-facial recognition mask
AI effect
Applications of artificial intelligence
Facial recognition system
Roads Policing Unit
Vehicle location data
Lists
List of emerging technologies
Outline of artificial intelligence
== References ==