Realising the potential of automatic number plate recognition

Emerging technology and big data analytics.

Summary

In recent years, the replacement of vehicle number plates to avoid detection has become a major policing issue. We refer here to those from any other vehicle as false plates, and to those from identical make, model and colour vehicles as clone plates. The UK’s Automatic Number Plate Recognition (ANPR) camera network offers the potential to explore the pervasiveness of this problem. This project sought to explore means to identify false and clone number plates from ANPR images using machine learning methods, and to thereby help realise the potential of police investment in ANPR infrastructure.

Key findings

• Machine learning proved successful at identifying false plates from ANPR images. It holds significant potential as a policing tool. Validation indicates our analytical workflow is 92% accurate in identifying false number plates. The tool that we developed is termed the False Plate Classifier (FPC).
  o In an unseen sample of 130,000 images, our analytics estimate the proportion of false plates where the vehicle make, model or colour differs to that registered, to be in the order of 0.03% (1 in 3700 vehicles).
• The FPC provides a proof-of-concept detection algorithm which, in its current format, is highly accurate in detecting false plates on vehicles with a different make, model or colour to that registered. Further model development and data sources are required to progress the detection of clone plates, but our results are highly encouraging.
  o We note also that the transfer of ANPR data from police to a university, in the context of GDPR, requires considerable commitment and technical skill from both parties but, as demonstrated here, is feasible. Data cleaning, handling and wrangling issues are complex, and require considerable investment. Machine learning requires specialist research expertise.
• The development of a software tool for police use, and an operational trial in a police control room, would be the appropriate next steps alongside continued development of detection methods.
• Further investment in ANPR-related research, and continued development of a False Plate Classifier, would help realise the potential of the significant and continuing investment in ANPR infrastructure.

Authors: Graham Farrell, Adam Hardy and Dan Birks (University of Leeds) and West Yorkshire Police – December 2019
Further information – Graham Farrell (G.Farrell@leeds.ac.uk)