A Support Vector Machines network for traffic sign recognition

The objective of this paper is to describe an algorithm able to solve the traffic sign recognition problem, based on a structure composed by a cascade of competing classifiers and some computer vision pre-processing operations. Traffic sign recognition is a very complex problem, involving a multiclass analysis with unbalanced class frequencies, most of them very similar to each other. With our system, that we are going to call Traffic Sign Classifier (TSC), during the competition promoted by the Institut für Neuroinformatik, Ruhr Universität Bochum, it was possible to recognize more than 40 classes of signs with an average error close to 3%. The algorithm, realized by our development team, consists basically of two modules: a preprocessing module, where the data are managed in order to extract some features, such as the Hue Histogram (HH) and the Histograms of Oriented Gradients (HOG); a second module, where the data coming from the first one are analyzed using a sequence of Support Vector Machines (SVM), implemented with the One Versus All (OVA) methodology. This module includes a couple of systems, composed of several SVMs; one of these systems consists of a hierarchical structure. The results coming out from both the systems are compared with each other in order to define which is the most reliable. This work is performed by the so called “Combining the Results and Assigning the Labels” procedure; calibrating the systems and the parameters employed inside the several analyses performed, it is possible to decrease the number of misclassifications and consequently increase the performance of the entire network.