Automated Traffic Sign Recognition System Using Computer Vision and Support Vector Machines

Abstract

This paper describes the initial design of a computer vision application to recognize regulatory traffic signs vertically installed on Colombian roads using machine learning. This application is conceived as a module of a driver assistance system under development, and an autonomous vehicle adapted to the local infrastructure. The application was trained and tested with official synthetic images provided by the National Ministry of Transport. These images were modified with chromatic and geometric changes to emulate fluctuations in illumination, vantage point, and ageing. Resulting images were resized to 48 × 48 pixels, and the raw intensity planes in the Hue-Saturation-Intensity color model were reshaped to obtain feature vectors with 2304 attributes each. In total, forty seven binary classifiers were trained using Support Vector Machines under a one-versus-all classification scheme. These classifiers were directly combined into a multi-class classification system. This paper reports the methodology used to collect the data, configure, train, and measure the performance of classifiers working isolated and collectively.