Focus of attention control

This report proposes an approach to control of spatial focus of attention using a structure for response generation where the cognitive part of a robotic vision system does not explicitly de ne every response made by the system. Instead it invokes general response commands that are modi ed by data, transforming them into situation speci c commands. In section 1, general aspects of response generation is discussed. Section 2, applies the the proposed ideas on the problem of controlling spatial focus of attention. Finally, section 3 contains the results from a series of simulations of spatial focus of attention. 1 Response generation structures Image Analysis Response Generation Figure 1: Classical pipelined structure of a robotic vision system. In robotic vision we often assume an image analysis system followed by a response generation system, see gure 1, [3]. The image analysis system is seen as an inputoutput system taking in images and producing a detailed description of, or at least hypotheses about, the world. The description is used by the response generation system to plan all movements in detail. The results of the response is then analyzed and further action is planned. There are several drawbacks with such a structure on a robotic vision system e.g. the sensor-motor path is rather long, which makes it hard for the system to cope with unexpected events. The close relationship between analysis and response mechanisms should be utilized especially in real world problems where responses are a natural part of information gathering. Reprinted from proceedings of the 7th Scandinavian Conference on Image Analysis, Aalborg, Denmark, August 13{16, 1991. 2 1.1 Analysis versus Response Instead of the classical approach with an analyzing system in one end and a responding system in the other, we would like to propose a di erent conceptual structure based on an integration of response and analysis. It is then important to consider that response mechanisms are an important part of analysis and vice versa. Levels of information abstraction Levels of response specification Sensor inputs Actuator outputs