A Unified Framework on Reading Variable Size Traffic Sign Boards using Deep Neural Network

Abstract

With the explosive increase of vehicles in modern era and continuous interest in autonomous vehicles across the globe, need for automatic signs reading software has attained global attention. Recent research has focused on reading small sized sign boards whereas previous work was more related to understanding large sized sign boards. This work serves twofold- identifying small or large sized sign boards within a single platform. This model makes use of recent advancements in deep learning paradigm entitle Inception V3, where this model is modified to cater needs of reading variable sizes of sign boards installed at traffic roads. This modified inception model is fed into Faster RCNN for better features extraction. RPN is applied for region of interest extraction. Later on, sign board is identified into three major classes, i.e. prohibited, warning or mandatory using fully connected layers of Neural Network. The proposed model achieves an accuracy of around 92.99% on a standard dataset.