High Accuracy Optical Flow Estimation Method Based On Sparse Model

Optic flow estimation is one of the basic problems in computer vision, and a lot of excellent optical flow estimation methods have been proposed. Recently, inspired by the development of compressed sensing, sparse prior is introduced as a novel regularization for optic flow estimation. This thesis focuses on the sparse prior of optical flow gradient field, and provides high accuracy optical flow estimation methods in three aspects:robust sparse signal recovery method, parametrized motion model for enhanced sparsity, and accurate explicit handing of occlusion.Firstly, based on the sparse prior of optical flow gradient field, we rigorously formu-late the optical flow estimation as an l0optimization problem. To facilitate the solution, according to the theory on sparse signal recovery, we relax the l0problem to a original l1problem, and reweighted scheme is employed to exploit the sparsity. Experiments show that the reweighted l1method leads to a more accurate sparse solution. In order to deal with discontinuities and noise effectively, we use the l1norm instead of the traditional l2norm constraint on the constant intensity constraint, then we have a more accurate optical, flow estimation result.In order to solve the problem that sparse optical flow field have poor performance in the motion regions with rotation and scaling, we generalize the sparsity regularized optic flow estimation to parametrized motion model. We show that the sparsity of the motion field can be enhanced by increasing the degree of freedom of the parametrized motion model, accompanied with the optical flow estimation getting more ill-posed. Experiments show that, with appropriate degrees of freedom parameter model, get the best optical flow estimation.Occlusion is one important factor affecting optic flow estimation. To deal with oc-clusion, we propose an effictive optic flow method based on occlusion conscious sparse model, and proveide a better description of the sparse optical flow estimation model. Dif-ferent to previous sparse models, this model handles occlusion explicitly. We detect the pixels related to occlusion, and initialize the optic flow of these occlusion pixels with directional patch matching. To achieve efficient solution, we relax the gradient sparse model to the total variational framework. We propose a reweight scheme to improve the estimation at motion boundaries. The reweight scheme introduces very slight additional computational load but greatly improves the estimation accuracy.This thesis estimate the optical flow based on the sparse model, and deal with three improtant problems:the sparse signal solution strategy, sparse prior selection and promo-tion, occlusion processing. We provide effective strategies to enhance the performence of optical flow estimation.