Representing navigational affordance based on high-level knowledge of scenes