Object localisation and tracking through subsymbolic classification

Abstract

This paper presents results of experiments in subsymbolic processing of visual data, to achieve identification and tracking of arbitrary objects, which are intended to be used in autonomous robots for novelty detection and navigation purposes. Artificial neural networks with unsupervised training are used as the classification stage for the vision system, in order to provide the robot the ability to develop its own representations from perceptual data without the need of any external human-provided information. We present an evaluation of the behaviour of the system when using very simple feature extraction techniques, such as horizontal and vertical average histograms, as well as average coarse coding.