Image Mapping and Visual Attention on a Sensory Ego-Sphere

Abstract

The research in this thesis focuses on two problems related to the Sensory Ego-Sphere (SES), a short-term memory structure for a robot: (1) the mapping to the SES of high-resolution sensory information in the form of imagery, and (2) the concurrent processing of visual attention. Neither problem had been studied previously. The SES coordinates sensory information for further processing and thereby acts as an interface between sensing and cognition. It is an egocentric, spherical mapping of the robot's locale. This research is based on previous work in the areas of multi-modal sensing, sensory-motor coordination, and attention. The paper describes a procedure to composite on the SES an image sequence taken by a camera head, a task that is complicated by significant overlap between successive images in the sequence. Two approaches to the problem of finding and ranking areas of visual interest are compared. One combines visual attention points in the overlapping images on the SES and the other computes attentional points directly on the composited visual scene. Computational structures for mapping imagery to the SES and managing it are described. The problem of attention in bio-vision is discussed as are some algorithms that mimic visual attention behaviors in humans.