Construction of Brazilian Regulatory Traffic Sign Recognition Dataset

Rafael Silva, Bruno O. Prado, Leonardo Nogueira Matos, Flávio Arthur O. Santos, Cleber Zanchettin, Paulo Novais

Abstract: In this article, we present the Brazilian Regulatory Traffic Sign Recognition Dataset, following the style of the CIFAR10 dataset. A convolutional neural network is also proposed to recognize and identify these traffic signs as a possible aid for ADAS (Advanced Driver Assistance Systems). The developed architecture has thirteen layers, selected after attempts to search for a sufficiently efficient organization. CNN used the RMSProp optimizer, a variant of the stochastic gradient descent technique (SGD), reaching 99.31% accuracy in training and 93.73% in the validation set. This document covers the dataset development process, convolutional neural network architecture, discussions about operation and results.