Ecuadorian regulatory traffic sign detection by using HOG features and ELM classifier

Abstract

This article presents an algorithm for Ecuadorian regulatory traffic signs detection, under extreme lighting conditions during the day. The method is composed of the following modules, i) video stabilization to reduce vertical oscilation, ii) a method to for obtaining regions of interest (ROIs) based on color information and geometric restrictions, iii) a two-stage multi classification algorithm, based on Extreme Learning Machine (ELM) and HOG descriptor, to classify by form and content. A database of Ecuador’s regulatory traffic signs has also been created. It has more than 47,000 images of Ecuadorian road signs divided into 16 classes. The experimental results, from the recognition module, generate an accuracy of 99,85 %, and a sensitivity of 99,78 % in the first stage, and an accuracy of 96,71 % and a sensitivity of 94,16 % in the second stage. In addition, its robustness was compared with two other classifiers, in order to choose the one with the best performance in terms of accuracy and low computational cost. This system works at 24 frames per second, and it was tested in real driving conditions, from 6:00 a.m. until 7:00 p.m., on the streets of several cities and highways of Ecuador.