Image-flow computation: Estimation-theoretic framework, unification and integration

Visual-motion is a major source of three-dimensional information. It is commonly recovered from time-varying imagery in the form of a two-dimensional image-flow field. This thesis is about computation of image-flow.; A new framework is proposed that classifies the image-flow information available in time-varying imagery into two categories--conservation information and neighborhood information. Each type of information is recovered in the form of an estimate accompanied by a covariance-matrix. Image-flow is then computed by fusing the two estimates using estimation-theoretic techniques. This framework is shown to allow estimation of certain types of discontinuous flow-fields without any a-priori knowledge about the location of discontinuities. In other words, the flow-fields estimated using this framework are not blurred at motion-discontinuities. Two algorithms based on this framework are described. Results of applying these algorithms to a variety of image-sequences are also discussed. In order to put the framework in context of an application, the image-flow fields recovered by these algorithms are used in a Kalman-filtering based approach to estimate the scene-depth.; The new framework is shown to be applicable identically to each one of the three major approaches for recovering conservation information, i.e., gradient-based approach, correlation-based approach and spatiotemporal energy-based approach. The formulation of neighborhood information used in this framework is also shown to reduce to some of the existing smoothing-based formulations under various simplifying assumptions. In essence, the framework described in this thesis unifies various existing approaches for image-flow computation. Such unification is useful in analyzing various existing frameworks as well as in generating new frameworks.; The new framework is also shown to serve as a platform to integrate the three approaches mentioned above. It is observed that the measurements obtained by the three approaches have different error-characteristics. This situation is regarded analogous to the multi-sensor fusion problem, where the algorithms based on the three approaches behave as multiple sensors measuring image-flow. An integrated framework is described that applies the principles of statistical estimation theory to fuse the measurements obtained from different approaches. The resulting estimate of image-flow has the minimum mean-squared error. Some algorithms based on this framework are described. Several proposals for extension of this thesis are also included.