Traffic sign recognition with convolutional neural network based on max pooling positions

Abstract

Recognition of traffic signs is vary important in many applications such as in self-driving car/driverless car, traffic mapping and traffic surveillance. Recently, deep learning models demonstrated prominent representation capacity, and achieved outstanding performance in traffic sign recognition. In this paper, we propose a traffic sign recognition system by applying convolutional neural network (CNN). In comparison with previous methods which usually use CNN as feature extractor and multi-layer perception (MLP) as classifier, we proposed max pooling positions (MPPs) as an effective discriminative feature to predict category labels. Through extensive experiments, MPPs demonstrates the ideal characteristics of small inter-class variance and large intra-class variance. Moreover, with the German Traffic Sign Recognition Benchmark (GTSRB), outstanding performance has been achieved by using MPPs.