Study on Optimal Convolutional Neural Networks Architecture for Traffic Sign Classification Using Augmented Dataset

Internet of Vehicles has become a very popular paradigm in the last years. Due to the technological progress and improved computing power, Deep Learning techniques are becoming a valid solution for multiple domains, including automotive. One area where Deep Learning algorithms are very useful is the traffic sign classification. In the last years, several solutions were proposed for this task, each of them having some downsides, usually related to the dataset quality or quantity. This paper presents a study on several architectures with a different number of layers, feature maps and input sizes with the goal of finding the best tradeoff for classifying traffic signs. Beside comparing architectures, this paper also presents a dataset augmentation method in order to avoid known dataset problems such as quality or quantity. The result of this study will be used in a second phase of real-time traffic objects detection.