Research On Stereo Matching Based On Super-pixel Segmentation

Stereo vision and 3D reconstruction technology are increasingly concerned in many field,such as VR,AR and MR.Stereo matching is the key of the 3D reconstruction so that it should be important to research and apply.However,the problem of occlusion and the optimization of non-submodule energy functions are still a major challenge.Currently there are many stereo matching algorithms such as the local SAD and NCC,the global belief propagation and graph cut.In this paper,we use the 3D labels to describe the spatial features of each pixel.Under the existing MRF-MAP energy cost model,the asymmetric graph model is modified to control the cluster size.All possible occlusion is presented in the graph by increasing the visibility nodes,and the related edges are integrated into the data items of the energy function.The second-order priori smoothing is used to relax the constraints of the parallel plane.The tangent planes are not penalized,so the pixels in the horizontal and vertical directions can be correlated with each other.Aiming to the occlusion and non-submodule function,we design an asymmetric graph model,while presents possible occlusion information in the graph by increasing the visibility nodes,integrating the weight of the visibility into data items.At the same time,QPBOI-R fusion move algorithm is introduced to optimize the non-submodule energy function.QPBOI-R creates another node which is opposite to each of the current pixel node,so that non-submodule edges can be segmented by min cut/max flow method.The approximate minimum cost solution can be obtained after times of iteration.This paper proposes a Multi-scale Super-pixel based Proposals(MSP)method.The multi-scale super-pixel image can be used to update the labels without affecting the depth changing at the edge.In addition,the combination of super-pixel proposals and two checkerboard proposals keeps the disparity image smoothing in the no-texture region and discontinued depth at the edge.