This research seeks to develop and apply recent advances in algorithmic methods for the construction of real time space-variant active vision systems. The term space-variant refers to sensor architectures in which pixel size, and possibly neighborhood topology, vary with position. Although inspired by biological vision (e.g. foveal vision, and its cortical expression in terms of the log-polar mapping) , space-variant architectures form a natural synergy with active vision: movement of the fixation point of a sensor, either via electronic pan-scroll, or via robotic actuators, is mandatory for space-variant systems. Since it has been shown that log-polar architecture can provide a form of data reduction that is on the order of two to three orders of magnitude for current computer vision systems, there is an important niche for application of space-variant imaging in robotic and computer vision. During the past year, an extension of the 2D Fourier Transform to log-polar image formats (the exponential chirp transform) has been developed. This transform has been demonstrated to provide the analog of a full-field (space-variant) 2D Fourier Transform at real-time rates on standard PC platforms. The goal of this research is to apply the exponential chirp transform to develop applications in the areas of real-time image tracking, template matching for recognition and motion de-blur, and to provide demonstrations indicating the achievement of complex visual functionality with minimal hardware and software complexity, for application on small, lightweight, and mobile active vision systems.
