On the Induction of Topological Maps from Sequences of Colour Histograms

Abstract

This paper presents an appearance­based method to automatically determine places from vision data for topological mapping. The approach exploits the continuity of the visual appearance of consecutive images when a robot traverses the environment. Places are determined by clustering colour histograms, and a probabilistic filtering strategy eliminates spurious places with weak evidence. Further, we discuss steps towards the induction of the topology of an environment from a sequence of visited places. Particularly, our system faces the problem of physically different places which appear identical in perception space. We present results from experiments on two data sets, one consist of panoramic images and another one includes images from a standard camera.