Research On Fast Small Baseline Stereo Matching Based On Image Segmentation

The problem of occlusion and geometric distortion of traditional stereovision in urban area elevation measurement can be resolved by small baseline stereovision.However calculation of disparity must reach sub-pixel level then make up the loss of elevation precision owing to minish the base-height ratio.Consequently,high precision sub-pixel stereo matching is a key problem in small baseline stereovision.This dissertation researches on the problem of small baseline stereo matching.A fast small baseline stereo matching based on image segmentation is proposed through analyzing the small baseline stereovision model.The small baseline stereo matching method is divided into three basic steps including integer pixel matching,disparity map refinement,sub-pixel matching.Consequently,the small baseline stereo matching is realized from coarse to fine according to the three basic steps.In order to improve the accuracy and precision of matching,a stereo matching framework based on image segmentation is adopted in this paper.In the integer pixel matching,the segmentation information is utilized to enhance the mutual support between pixels in the same region.In the disparity map refinement,the segmentation information is utilized to fit the disparity plane model for the textureless region.In the sub-pixel matching,the segmentation information is utilized to adjust the phase correlation window adaptively.The main research results of this dissertation are as follows:(1)A multiscale stereo matching method based on segmented cross tree and fuzzy logic is proposed to resolve the problem of integer pixel matching for small baseline stereoscopic image pairs.Firstly,the multiscale image pyramid of the stereo pair is constructed by the downsampling method.Secondly,the original image is converted from RGB color mode to HSL color mode,and the initial matching cost is calculated by using the matching cost function based on fuzzy logic in the top scale space,then the matching cost is aggregated by segmented cross tree.Thirdly,the disparity in current scale space is calculated by the “winner takes all” strategy.Finally,the final whole disparity map is calculated from coarse to fine under the guidance of the disparity of the upper scale space.The accurate integer pixel disparity is calculated for high precision sub-pixel small baseline stereo matching through this method which is efficiency and accuracy,and is robust to illumination variation.(2)A disparity map refinement method based on multilevel image segmentation is proposed to resolve the mismatch problem of textureless region in disparity map.Firstly,the multilevel image segmentation technique is used to divide the weak texture region in the stereo image into a series of non overlapping image regions.Secondly,the initial disparity is calculated for each region.Thirdly,the disparity plane model of each region is fitting based on the initial disparity.Finally,the adjacent disparity planes with higher similarity are merged.The textureless regions in disparity map is recognized and fitting as disparity plane model through this method which is improving the quality of the disparity map of textureless region.(3)A sub-pixel stereo matching method based on image segmentation and phase correlation is proposed to resolve the problem of high precision sub-pixel matching for small baseline stereo pair.Firstly,the relationship of corresponding points is established under the guidance of integer pixel disparity.Secondly,the phase correlation window is selected according to the segmented domains adaptively which takes the corresponding point as the center.Thirdly,the sub-pixel shifting is obtained by matching the windows utilizing an phase correlation matching method.Finally,the final sub-pixel disparity result is obtained by adding the integer pixel disparity and the sub-pixel shifting.The size of the phase correlation window is adjusted adaptively,and the influence of the local weak texture characteristics on the phase correlation peak is avoided through this method which is precision and efficiency.This dissertation mainly studies fast small baseline stereo matching method based on image segmentation.Three basic steps including integer pixel matching,disparity map refinement,sub-pixel matching are designed and implemented,which are utilized to match the stereoscopic image pairs.This method has high precision of the disparity map and is an effective and accurate small baseline stereo matching scheme.