Image segmentation for appearance-based self-localisation

Abstract

The paper describes a segmentation technique that well fits to an appearance-based self-localisation. In an appearance-based approach robot positioning is performed without using explicit object models. The choice of the representation of image appearances is fundamental. We use image-domain features, as opposed to interpreted characteristics of the scene, and we adopt feature vectors including both the chromatic attributes of colour sets and their mutual spatial relationships. To obtain the colour sets we perform image segmentation by autothresholding the colour histograms and taking into account what the results are addressed to. The experimental results indicate that the method performs well for a variety of environments.