Stereo active vision and peripheral optical flow: Computer vision applications of the wide-field human visual representation

The topographic structure of the central 20° of the primate visual field is approximated by a two-dimensional (complex variable) logarithmic function, a monopole map. Recent work shows that a second logarithmic pole in the far peripheral field provides a wide-angle model of visual representation. This dipole map is in good agreement with physiological data from the foveal to the far peripheral representation. The foveal and peripheral logarithmic poles have natural roles to play in object discrimination and egomotion, respectively. The first part of this work applies the foveal and parafoveal representation to the problem of designing a real-time stereo active vision system. A stereo “robotic head” was constructed with two cameras and actuators providing pan, tilt, and vergence. A set of attentional operators was developed for this system, which utilizes cues based on motion, color, depth, and shape, first for conventional digital images and then for monopole-mapped images. These attentional operators are demonstrated for the task of locating human faces in live video imagery. Next, optical flow near the peripheral logarithmic pole is used to construct robust navigational cues to sensor velocity. Most optical flow algorithms attempt to locate both the focus of expansion (FOE) and the axis of rotation (AOR) of the flow field, in order to estimate sensor heading and velocity. This approach generally fails to work in practice because it is highly sensitive to noise, motion distractors, and sensor platform jitter. However, this thesis shows that the peripheral flow field allows a robust extraction of sensor velocity without detailed knowledge of the FOE or AOR, and in the presence of significant unknown motion distractors. This thesis represents the first algorithms that directly exploit the wide-angle complex dipole model of human visual anatomy in computer vision applications.