Traffic Sign Identification Using Deep Learning

Abstract

One of the most crucial enabling technologies for automated driving systems is the ability to reliably detect and classify a wide range of traffic signs in various driving conditions at different distances. Due to the complexity and dynamic nature of driving environments, it is difficult to reliably detect traffic signs with conventional image processing methods. Artificial intelligence in combination with image processing has proven to be a great success to address this problem in recent studies. This paper focuses on the selection of Deep Neural Networks (DNN) based on the application-oriented performance by taking into consideration the mean Average Precision (mAP) and Frames Per Second (FPS) as the major evaluation criteria. Faster Region-based Convolutional Neural Network (Faster R-CNN) is a newly proposed DNN in the literature that has proven to exhibit a balanced tradeoff between mAP and FPS performance measures. This paper starts with a DNN transfer learning and then implements the Faster R-CNN algorithm for the real-time detection and classification of traffic signs using the Robot Operating System (ROS). To reduce the errors due to DNN inaccurate detection, Tesseract" is added to detect the text in the identified traffic signs. The German Traffic Sign Detection Benchmark (GTSDB) dataset is used in this paper, and additional dataset are created to solve the lack of certain traffic signs in the GTSDB dataset. Simulation with ROS-Gazebo and real-time trials using the Polaris Gem e2 equipped with NVIDIA Drive PX2 demonstrate the efficiency of the proposed integration of DNN with Tesseract in detecting and classifying a wide range of traffic signs.