Research On Regularization Term In Variational Optic Flow Computation For Motion Segmentation

Optic flow is a key technique for motion image analysis. Variational methods are among the best performing techniques for computing the optic flow. Such methods determine the optic flow field as the minimiser of a suitable energy functional, which consists of two terms:a data term that imposes constancy on image features, and a regularization term that induces a smoothness constraint. The accuracy of optic flow estimate would be improved significantly if adaptive regularization term could be induced. This thesis focuses on the regularization term design and meaningful attempts have been done in the application of optic flow for motion segmentation. Our research works include:●Due to the theory of diffusion is the foundation of regularization term design, an investigation on diffusion theory has been carried out. Then different regularization terms are thereby classified in accordance with the induced diffusion process. The advantages and disadvantages of each class are investigated deeply at the same time.●By analyzing the existing design methods, a new one is proposed which design regularization term by joint image-& flow-driven. The study proved that this proposed approach can effectively overcome the deficiencies of the other methods and preserve the discontinuities in optic flow field more accurately.●Utilizing the proposed regularization term and integrating the non-linearized constant assumption, an advanced optic flow model has been developed. And we also present an efficient numerical scheme for solving the model. The results of experiments show that this method has specific advantages for the complex scenes.●The traditional active contour model has been improved to make it more effectively for motion segmentation. At the same time, combined the high quality optic flow information which is acquired from the improved model, a motion segmentation algorithm are presented. Then the experiments demonstrated the validity of this algorithm.Drawn by the requirements of the high quality optical flow field, the problems of regularization term design and the corresponding numerical scheme are studied in this thesis. An advanced regularization term for optic flow estimate has been proposed. The work of the thesis is meaningful for optic flow estimate and motion segmentation.