Research On Accurate Stereo Matching Algorithms In Stereo Vision

Stereo matching is one of the critical issues in stereo vision for depth perception and its main aim of stereo matching is to obtain accurate information of 3D scenes.At the same time,because stereo matching performs insensitive to illumination and more robust in outdoor environment,it has been widely applied into 3D Reconstruction,Image Refocusing,View Interpolation and so on.But it still possesses some problems,such as lack of 2D and 3D information combination,easyily mismatching in textureless regions,high computational complexity and being unable to effectively use of multi-scale information,which limit the further improvement of the accuracy.For solving the above mentioned problems,this thesis does the in-depth research into four sections as follows:(1)The research on tree based non-local cost aggregation stereo matching algorithms.For solving the problem of combining 2D information with 3D information,an Iterative Color-Depth Minimum Spanning Tree cost aggregation stereo matching algorithm is firstly proposed.For this algorithm,3D information is introduced to construct the minimum spanning tree structure,and the textureless sensitive matching cost computation method is given to improve the problem of mismatching in textureless regions.Then,an Improved Segment-Tree based Cost Aggregation stereo matching algorithm is proposed.When constructing the segment-tree structure,the algorithm not only introduces 3D information and textureless sensitive matching cost computation method,but also improves the segmentation strategy to make it more consistent with disparity consistency assumption.The experimental results demonstrate that the accuracy of the above algorithms is better than other homogeneous algorithms in integer accuracy while maintaining real-time performance.(2)Spatial PatchMatch Stereo Matching Algorithm with Virtual Pixel Aggregation.This algorithm is based on the Nearest Neighbor Field(NNF),the step of “View Propagation” is firstly removed to reduce the error disparities.Then a method of four neighbored Spatial Propagation is utilized to further improve the accuracy of the algorithm.Lastly,a Virtual Pixel Aggregation strategy is combined to effectively reduce the mismatching rate in textureless regions.Experimental results demonstrate that the proposed algorithm could yield more accurate results than other homogeneous algorithms in subpixel level accuracy.(3)The Global Algorithms Taking Cost Aggregation as Data Terms of Energy Function.In order to effectively improve the computational complexity of stereo matching algorithm,an Accelerating Global Optimization for Accurate Stereo Matching algorithm is firstly proposed.The algorithm takes the non-local cost aggregation as data term of energy function,and the constraint of two neighbored pixels as the smoothness term to optimize the whole energy function.On this basis,the optimization efficiency of the algorithm is promoted by improving the optimizer.Then,an algorithm,MeshStereo with Cross-Scale Cost Filtering for Fast Stereo Matching is proposed.The algorithm takes the Cross-Scale Cost Filtering model as data term of energy function of the MeshStereo model,which solves the problem that the multi-scale information unable to effectively utilized.At the same time,the computational complexity of the algorithm is reduced by using the MeshStereo model.Finally,the normal and depth constraints are used as smoothness terms to optimize the whole energy function.Experimental results demonstrate that both the global stereo matching algorithms can obtain more accurate disparity estimation.(4)As-Global-As-Possible Stereo Matching Algorithm with Adaptive Smoothness Prior.This algorithm transforms the 1D scanline optimization into a 2D optimization opinion,also utilizes the similarity of neighbored pixels as smoothness prior,carries out cost aggregation based on energy function,and reduces the computing time of the algorithm through multi-CPU parallel processing.Compared with the homogeneous algorithms,the proposed As-Global-As-Possible Stereo Matching Algorithm with Adaptive Smoothness Prior has higher accuracy of matching.