Selection of an Optimal Set of Discriminative and Robust Local Features with Application to Traffic Sign Recognition

Abstract

Today, discriminative local features are widely used in different fields of computer vision. Due to their strengths, discriminative local features were recently applied to the problem of traffic sign recognition (TSR). First of all, we discuss how discriminative local features are applied to TSR and which problems arise in this specific domain. Since TSR has to cope with highly structured and symmetrical objects, which are often captured at low resolution, only a small number of features can be matched correctly. To alleviate these issues, we provide an approach for the selection of discriminative and robust features to increase the matching performance by speed, recall, and precision. Contrary to recent techniques that solely rely on density estimation in feature space to select highly discriminative features, we additionally address the question of features’ retrievability and positional stability under scale changes as well as their reliability to viewpoint variations. Finally, we combine the proposed methods to obtain a small set of robust features that have excellent matching properties.