Neural Network-based Multi-Class Traffic-Sign Classification with the German Traffic Sign Recognition Benchmark
Csanád Ferencz, Máté Zöldy
Abstract
Traffic-sign detection has an essential role in the field of computer vision, having many real-world applications more and more object recognition and classification task is being solved by using Convolutional Neural Networks (CNNs or ConvNets), especially in the field of intelligent transportation. In the present article, we offer an implementation chosen from several CNN-based traffic-sign recognition and classification algorithm architectures, using a ConvNet classifying 43 different types of road traffic signs in the TensorFlow framework, as part of the German Traffic Sign Recognition Benchmark (GTSRB) competition. A Deep ConvNet was trained end-to-end, aiming to improve the prediction performance of a DCNN-based autonomous driving system equipped with a front-facing digital camera, with as input a sequence of images, as output directly the prediction results.The results obtained on held-out data demonstrated the high accuracy of the model, matching the state-of-the-art multi-class recognition and classification accuracies, as well as related human-level recognition performances.