Deep Learning with Accelerated Execution: A Real-Time License Plate Localisation System

Abstract

In this paper, a real-time, real-life, novel license plate localisation (LPL) based on deep learning (DL) and accelerated by field-programmable gate array (FPGA), and Open Visual Inference and Neural network Optimization (OpenVINO) toolkit, is proposed and prototyped. The solution was tested against two popular international research databases and achieves state-of- the-art results, proving the viability of FPGA in real-life latency- oriented application. Using novel asynchronized DL inference that prepares next result while current inference is ongoing, the system increases computational efficiency without buffering frames, allowing for reduced latency. Comparisons show that the proposed LPL system has lower latency and better performance per watt than other related solutions.