This thesis concerns the early processing of visual motion in biological and machine vision systems. We are particularly interested in the phenomena of non-Fourier motion, a topic of significant recent interest in the perception literature. This thesis investigates a model of non-Fourier motion, which is based, in part, on the orientation of power in frequency space. Toward this end we construct an algorithm in which the amplitude gradient of the output of band-pass filter is used to extract non-Fourier motion. We unite the mechanism for both Fourier and non-Fourier motion with the use of quadrature filters, the phase of which provides information about Fourier motion and the amplitude provides information about non-Fourier motion. In the extraction of the non-Fourier motion we utilize a low-pass average operation in the spatial domain, which is equivalent to finding the optimal orientation of the spectral power in a Least-Mean-Square sense. The algorithm is tested on a significant collection of test cases for which we present both qualitative and quantitative results. This is one necessary step in creating a type of algorithm that is basically consistent with biological findings, but is computationally capable of extracting the relevant sources of information.
展开▼