Abstraction, ontology and task-guidance for visual perception in robots

Abstract

For solving recognition tasks in order to navigate in unknown environments
and to manipulate objects, humans seem to use at least the following crucial 
capabilities: abstraction (for storing higher-level concepts of things), common 
sense knowledge and prediction. Whereas the first and second provide the 
basis for situated recognition, the second and third serve for pruning the 
search space as it helps anticipating what (in an abstract sense) they will see
 next and where. The main goal of our current research is, how we could use 
such a kind of "common sense world knowledge" for guiding visual perception
 and understanding scenes. Therefore, we are combining an owl-ontology 
with the output of vision tools. The additional use of abstraction techniques 
tries to establish the possibility of detecting higher level concepts, such as 
arches composed of a variable number of parts. The goal is to finally find 
concepts such as doors and tables in arbitrary scenes in order to arrive at a 
generic recognition tool for home robots. The ontology should additionally 
provide task-specific information about the things to detect.