Piecewise Rigid Scene Flow Estimation Based On Edge Preserving Interpolation

With the development of computer technology and artificial intelligence technology,computer vision has become a hot area of research.The estimation of scene flow through a stereo image sequence is an important research content in the field of computer vision.Scene flow estimation is the main means of 3D motion research.Scene flow represents the 3D motion field of the real world,which provides more reliable information for motion estimation.With the continuous improvement of computer software and hardware,many new algorithms of scene flow estimation are emerging constantly,which promote the application of scene flow technology in real life.For example,scene flow is applied to driverless,3D reconstruction,target detection and tracking.Scene flow estimation is still affected by the problem occluded and edge motion not continuous and so on.In this paper,the problem of scene flow estimation is optimized to further improve the estimation accuracy of scene flow.Using the information obtained from the binocular stereo image sequences,constructing the energy functional of the variational scene flow estimation based on the piecewise rigid assumption,and then a piecewise rigid scene flow estimation algorithm based on edge preserving interpolation is proposed.First,constructing the energy functional of the scene flow estimation based on the piecewise rigid assumption.The energy functional consists of two parts of the data term and the smoothing term.The data term based on the brightness constancy assumption constraint,and the robust penalty function is used to enhance the robustness of the functional.The smoothing term includes space regularization term,piecewise regularization term and visiblility regularization term.The classical scene flow estimation techniques mostly use the global pixel level smooth constraints,which not only increase the amount of calculation but also be easily restricted by the underlying pixels.In this paper,a piecewise regularization term constraint is used to transform the global pixel level smooth constraint into piecewise plane smooth constraint.It is assumed that the object has a uniform motion in a segmented region.In the process of solving,the piecewise plane is continuously optimized.A smooth constraint based on piecewise rigid can not only greatly reduce the amount of data calculation,but also improve the accuracy of scene flow estimation.The scene flow estimation can produce large error in the motion edge part.For this problem,this paper use the reverse cross consistency check to exclude the error area to get the sparse scene flow,and then use the edge preserving interpolation method to interpolate the sparse scene flow into dense scene flow.The sparse to dense interpolation depends on the distance between the interpolated point and the other valued points.The calculation of the geodesic distance is based on the edge of the image,so the quality of the edge detection graph has a great influence on the interpolation.The edge detection part of this paper uses a fast edge detection structured forest(SED)algorithm,which can achieve better edge detection.Finally,the validity of the algorithm is verified by using the Teddy and Cones data sets on the Middlebury website,the Kitti dataset and the real scene image sequences acquired by binocular stereo camera.