Vision for situated robot companions — Fusing top-down knowledge and bottom-up data

Abstract

In this paper we try to countersteer the observed dissociation of computer vision research from robotics and artificial intelligence. We propose a theoretical framework of cognitive functions with which the overall agent's knowledge (serving as top-down information repository) and bottom-up vision data can be glued together. It is argued that vision is always intentionally directed, working in a concrete situation for a concrete task. We call this paradigm situated vision and argue for a variety of different vision techniques working with and on the same ontology of the agent. We present three distinct methods and show how they may support a robot in a concrete showcase example.