Automatic Place Determination using Colour Histograms and Self-Organising Maps

Abstract

In this paper we propose a model-free appearance-based method to automatically determine places as landmarks for topological navigation. Most of the current approaches for automatically selecting landmarks are based on template models or complex feature detectors. We use modified colour histograms, more precisely an entropy-constrained 3D~colour clustering as appearance-based image features which adapt to the colour distribution of the environment.
An unsupervised neural network learning strategy is used to automatically determine places by clustering the modified histograms.
Results from experiments in an indoor environment with a robot equipped with a panoramic camera show that the places, which were clustered in histogram space refer to physically close positions in the world domain in a large degree and can be used as landmarks for navigation purposes.