Research On Accurate Stereo Matching Algorithms Based On Image Filtering

Stereo vision is a research hotspot and frontier problem in the field of machine vision.Compared with the active deep acquisition methods based on the structured light or the time of flight technologies,the stereo vision has the advantages of passive depth acquisition,high resolution,low power consumption and low cost,and is widely used in aerospace,remote sensing,and automatic driving.In binocular stereo vision,the stereo matching algorithms are used to obtain the disparity maps of various scenes that can be further applied to calculate the depth information.Among the current binocular stereo matching algorithms,the image-filtering based local methods are of low computational complexity and high operational efficiency.It has become the focus of academic research and industry research,but there are still problems in proper selection of window size,poor matching accuracy in texture-less regions,as well as restricted refinement performance in small disparity regions.To solve or alleviate these problems,this thesis has carried out the following works and proposed several effective stereo matching approaches:1.Aiming at solving the window selection problem in the traditional stereo matching algorithm,an algorithm based on the permeability filer and the non-local weighted guided image filtering are proposed.Weights of the algorithm incorporate both spatial and brightness differences,which include a step function to improve the weight attenuation phenomenon common to the non-local filtering.Experimental results using the Middlebury version 3 dataset show that,compared with the state-of-the-art algorithms based on guided-image filtering,the proposed algorithm eliminates the effects due to the size selection of support window.The weighted average matching errors in the non-occluded and all regions are reduced by 1.81% and 1.25%,respectively.2.For the matching difficulties in texture-less regions that troubled many advanced guided-filter matching algorithms,this thesis further extends the non-local permeability filter to be integrated in a multi-scale scheme that takes features in different resolutions into account.This effective integration of the parameters of the guided filter models at different scales improves the matching accuracy in weak texture regions.The corresponding experimental results,also using the Middlebury version 3 benchmark dataset,show that the algorithm can effectively improve the stereo matching accuracy in the weak texture regions.The weighted average matching errors in the non-occluded and all regions are further reduced by 1.72 % and 1.78 %,respectively.3.Disparity refinement is an important step to enhance the accuracy of stereo matching.However,traditional cost aggregation methods based on the minimum spanning tree(MST)suffer from limited refinement performance in small disparity regions.A new disparity refinement scheme is proposed that exploits the initial disparity map for the estimation of disparity difference.The scheme embraces a constraint modification to the initial matching cost.Performance verification is ensured by extensive experimental comparisons,again,using the Middlebury Stereo Evaluation version 3 datasets.Accuracy of the scheme outperforms the MST-based disparity refinement approaches in all regions.The weighted average matching errors in the non-occluded and all regions are improved by 1.20 % and 1.09 %,respectively.