An empirical study on traffic sign recognition using convolutional neural networks

Abstract

This thesis presents a system for recognizing and classifying road and traffic signs. The purpose of this system is to develop an inventory of these signs, which can assist highway engineers in updating and maintaining them. Unveiling traffic signs (TSD), the main technology of TSR, is a significant issue due to various styles, small size, challenging riding conditions and obstacles. At first, images were collected from online sources using devices such as cameras, mobile phones and electronic devices from a moving vehicle. Then it was pre-processed by resizing and converting to grayscale. This thesis is based on CNNs. As CNN requires input images to be a certain size and in grayscale, they must be pre-processed before being fed into the CNN module. In the CNN module, the present researcher uses three convolution layers with 32 filters, 64 filters and 128 filters, respectively. Also, it has three max pooling layers and a ReLU activation layer. Dropout layer of 0.5 is present to prevent the occurrence of overfitting problem. Here in this thesis, the present researcher has utilized a substantial amount of data for both training and testing purposes. The present researcher has used a total of 39,209 images for training and 10,000 images for testing purposes. The present researcher has successfully achieved a 99.52% accuracy rate, which is an improvement compared to the previous work on Traffic Sign Recognition using Convolutional Neural Network.