A Fast Algorithm For Large Displacement Optical Flow Calculation Based On Variational Method

The so-called “optical flow",meaning restoration of displacement field from two continuous frames of image sequences,can become the basis of image understanding,image analysis and many other related fields.How to estimate the optical flow is a key problem in computer vision field.The variational method considering problems of optical flow computation from mathematics for optical flow computation provides a new and perfect theory frame.After decades of development,the variational technique for optical flow has been a very mature measure,but the efficiency problem of the optical flow computation has been bothering people.The optical flow computation can be divided into small d isplacement,large displacement.In the domain of small displacement optical flow computation,we summarize for several iterative methods for solving variational model,and proposed a simplified algorithm based on the principle of variational optical flow model,and the simplified algorithm greatly reduces the computation cost.In the domain of large displacement optical flow computation,the multiple scales method make small displacement optical flow models calculate large displacement optical flow,but increased the burden of calculation of the variational model.In this paper,we proposed three fast methods of large displacement variational optical flow model to solve the problem in the process of solving non-differentiable problem and to speed up the convergence rates which can improve the computational efficiency.Experiments show that three methods can improve the computational efficiency in ensuring the accuracy of the premise.The edge preserving ability of isotropic diffusion model is generally weak.In this paper,the regularization term of L1 norm total variation model is applied into the optical flow model to propose a new nolinear diffusion model for large displacement optical flow model and a fast method is used for the new model.Experimental results show that the proposed model has good edge preserving ability.