Integrating Aspects of Active Vision into a Knowledge-Based System

While the strategy of active vision is well established in early vision, it is not widespread in high-level vision. In this paper we suggest an approach for integrating aspects of active vision into a knowledge-based system. One aspect is the selection of optimal camera actions which are chosen to make the recognition process more reliable and efficient. We integrated such camera actions into our knowledge base. In addition, we describe the extensions of the control algorithm which is needed to use the information represented in the knowledge base, closing the loop between acting and sensing. Experiments show the efficiency and flexibility of the system. As an example, the task of locating objects in an office room is evaluated.