Localizing License Plates in Real Time with RetinaNet Object Detector

Abstract

Automatic license plate recognition systems have various applications in intelligent automated transportation systems and thus have been frequently researched over the past years. Yet designing a highly accurate license plate recognition pipeline is challenging in an unconstrained environment, where difficulties arise from variations in photographic conditions like illumination, distortion, blurring, etc., and license plate structural variations like background, text font, size, and color across different countries. In this paper, we tackle the problem of license plate detection and propose a novel approach based on localization of the license plates with prior vehicle detection, using the state-of-the-art RetinaNet object detector. This helps us to achieve real-time detection performance, while having superior localization accuracy compared to other state-of-the-art object detectors. Our system proved to be robust to all those variations that can occur in an unconstrained environment and outperformed other state-of-the-art license plate detection systems to the best of our knowledge.