Information Sampling for vision-based robot navigation

This paper proposes a statistical, non-feature based, attention mechanism for a mobile robot, termed Information Sampling. The selected data may be a single pixel or a number scattered throughout an image. After ranking this data, we choose only the most discriminating to build a topological representation of the environment, obtained via Principal Component Analysis (PCA). Advantageously, using this approach, our robot gains the ability to make effective use of its perceptual capabilities and limited computational resources. Real world experimental results verify that vision-based navigation is possible using only a small number of discriminating image pixels. © 2002 Elsevier Science B.V. All rights reserved.