Targeting a Practical Approach for Robot Vision with Ensembles of Visual Features

Abstract

We approach the task of topological localization in mobile robotics without using a temporal continuity of the sequences of images.  The provided information about the environment is contained in images taken with a perspective colour camera mounted on a robot platform.  The main contributions of this work are quantifiable examinations of a wide variety of different global and local invariant features, and different distance measures. We focus on finding the optimal set of features and a deepened analysis was carried out. The characteristics of different features were analysed using widely known dissimilarity measures and graphical views of the overall performances. The quality of the acquired configurations is also tested in the localization stage by means of location recognition in the Robot Vision task, by participating at the ImageCLEF International Evaluation Campaign. The long term goal of this project is to develop integrated, stand alone capabilities for real-time topological localization in varying illumination conditions and over longer routes.