Real-Time Classification of Traffic Signs with Deep Learning

Abstract

Problems are commonly encountered in image classification tasks within the field of computer vision. Over the past few decades, such problems have increasingly been treated with machine-learning techniques and, in particular, with deep learning and artificial neural networks. Due to the recent rapid development of Graphics Processing Units, sufficient computational resources needed for training deep neural network models with large datasets have become available. This thesis uses the methods of deep learning to investigate an image classification task where the input data consists of photographic images of various different traffic signs. The aim of the thesis was to build an automatic classifier that is able to recognize the type of a particular traffic sign automatically from its image. For this purpose, a convolutional neural network model was created and trained with a large dataset containing tens of thousands of images of over 40 different traffic signs. After training, the final classifier was incorporated into a computer program designed for detecting traffic signs in real time with the help of a webcam. The project was conducted with the Python programming language using the Keras framework and AWS SageMaker hardware. The convolutional neural network used in this project is a customized version of the LeNet-5 model architecture. The outcome of the project is a combined system consisting of a webcam and a classifier that is able to recognize and predict the type of traffic signs from their images in real time and with a high rate of success.