Development of a Real-Time Traffic Sign Recognition System Based On Deep Learning Approach with Convolutional Neural Networks and Integrating to The Embedded Systems

Akgul B. A. , Haznedar B. , Hasoğlu M. F. , Bayram K. S.

Euroasia Journal of Mathematics Engineering Natural and Medical Sciences, vol.8, no.14, pp.46-64, 2021 (International Refereed University Journal)

  • Publication Type: Article / Article
  • Volume: 8 Issue: 14
  • Publication Date: 2021
  • Doi Number: 10.38065/euroasiaorg.481
  • Title of Journal : Euroasia Journal of Mathematics Engineering Natural and Medical Sciences
  • Page Numbers: pp.46-64


Traffic signs are mandatory features of road traffic regulations worldwide. Automatic detection and recognition of traffic signs by vehicles may increase the safety level of drivers and passengers. For this reason, Real Time-Traffic Sign Recognition (RT-TSR) system is one of the essential components for smart transportation systems and high-tech vehicles. Recently, very good performances have been achieved in public datasets, especially with advanced Computer Vision (CV) approaches like Deep Learning (DL). Nevertheless, these CV techniques still need to be improved to provide the requirements of Real-Time (RT) applications. Although hopefully outcomes have been obtained theoretically in previous Traffic Sign Recognition (TSR) studies, there are very few studies that offer RT solutions in the real world. Therefore, in this study, a DL-based RT-TSR system is developed because of its high rate of recognition and quick execution. Besides, the CV approach has been included in the software developed to achieve classification as RT and support digital imaging. This developed system is capable of running smoothly in embedded systems with mobile GPU or CPU thanks to its low computational cost and high performance. Therefore, this study makes two important contributions to this field: software and hardware. First, RT-TSR software has been developed by using Convolutional Neural Networks (CNN) built on DL techniques along with CV techniques. Secondly, the developed software is adapted to embedded devices and hardware design is made. This developed system is also a technology product that offers software and hardware solutions together. Coding is accomplished under TensorFlow and OpenCV frameworks with the python programming language and CNN training is carried out by using parallel architecture. The experimental findings indicate that the developed CNN architecture achieves 99,71% accuracy and confirms the high efficiency of the system.