Designing the ANPR Pipeline using State of the Art Techniques

Abstract

Automated Number plate recognition (ANPR) is a subject area undergoing intense research and adoption in recent times. Various combinations of state of the art techniques and methodologies have been proposed to solve the problem end to end, with mixed results. These techniques include computer vision based mathematical notations and training based on artificial neural network (ANN). Recently ANPR has seen adoption in areas such as parking optimization, security surveillance in public and private places, law enforcement and traffic violation tracking. Selecting the most efficient and accurate algorithm and method depends heavily on the underlying use case being addressed as well as on prevalent environmental factors. This is because the intrinsic factors and challenges of a particular use case can translate into many subtle changes to the input frame. For example, In India, vehicle license plates can have characters of different languages, fonts and styles. And an implementation for a parking area may be subject to different conditions of lighting when compared to one for traffic signals or security surveillance. As such it is extremely hard to find a one-size fits-all solution to this problem. In this paper, we will discuss some of the current methods such as Morphology, LBP classification, OCR, CNN and others. We will highlight the results we obtained when using each of them and some of their pros and cons when applied in certain conditions. This will assist in building an efficient ANPR pipeline architecture starting from live feeds and ending at real time actions.