Scene Flow Estimation Based On Adaptive Anisotropic Total Variation Flow-driven

The world we live in is complex and changeable.In the process of exploring the world,most of the information comes from the visual.With the rapid development of computer technology and sensor,machine vision becomes an important part of automatic data acquisition.In the visual field,the moving object contains more information than the stationary object and the moving flow field becomes the bridge between the bottom information to the high-level image analysis.2 dimensional optical flow describes the change of the brightness pattern of the moving object surface.The scene flow describes the 3dimensional dynamic scene and can be regarded as the extension of optical flow in the 3dimensional space.The estimation of flow field is an ill-posed problem and its solution needs to add a variety of assumptions in the variational framework to solve the regularization problem.Because the variational model can transform the flow field into extreme value problem,it shows great advantage in estimation of the flow field.This paper gets information from stereo image sequences and construct the scene flow estimation variational model.By combining with the image sequences in time and space,data terms combine the brightness and gradient constancy assumption and enhance algorithm robustness.For each components of data term,occlusion factor is adopted to deal with occlusion problem.Smoothing term is an important part of solving the flow field and achieve constraint for the model by imposing penalty on variables.Data terms with image sequence of time and space combine the brightness and gradient constancy assumption,improving the robustness of the algorithm.Each component in the data term handle the occlusion problems with shielding factor weighting method.Smoothing term is an important part of solving the flow field by applying the penalty to the variables of the method.This paper presents an adaptive anisotropic flow driven scene flow estimation method.This paper analyzes the isotropic flow driven smoothness which has the same diffusion coefficient in each component and lacks direction.On this basis,the anisotropic flow driven is proposed to improve the estimation accuracy and reliability of the moving edges.Because of the "staircase" in the flow field of the total variation model,the adaptive anisotropic flow driven model is proposed by adaptive weighting.For the large displacement problem in scene flow model,this paper introduce hierarchical refinement which transmit the values in coarse resolution layer to the fine resolution layer.After the first frame image is wapped,we then calculate the residuals of the latter frame image to overcomethe large displacement problem.In each layer,scene flow is solved by successive over-relaxation(SOR).Finally,the validity of the algorithm is verified by using the Cones,Venus and Teddy data sets on the Middlebury website,the hemi-spheres dataset and the real scene image sequences acquired indoors.