The Binocular Vision-based Stereo Matching Methods

Stereo vision theory aims at obtaining the depth information of the view scene,and it is closely related with several applications,such as 3D measurement,3D reconstruction,robotic vision,robotic navigation,automatic driving and so on.Being the most principal stage of stereo vision,stereo matching establishes correspondences between pixels in images of two viewpoints,which influence the accuracy of the depth information directly.With the development of stereo matching,the local matching method becomes the emphasis topic,which realizes real-time implementation.The major factors,which influence the effect of local matching methods,can be classified into two components: external factors,like noise,illumination variation,camera distortion;inherent factors,like support window selection,support weight calculation and cost aggregation.For the problem of camera distortion minimization,we propose an efficient three-step rectification method based on the epipolar distance transform measurement,which minimizes the relative rotation between both cameras step by step to limit the camera distortion,laying the foundation for the research of stereo matching.For the mismatches within the texture less regions and discontinuous regions,we provide a ‘butterfly-shaped' adaptive window selection method,a histogram-based adaptive weights aggregation method and an adaptive damping factors-based permeability filtering weights aggregation method,which improve the matching accuracy of local method.Considering the effects of noise and illumination variation,we raise an adaptive epipolar distance transform,which converts the grey values to a distance measurement,and match on image structure,the proposed distance measurement is robust to noise,and distinguishes the pixels within the texture less regions.The main work and innovation of this dissertation are presented as follows:1)Considering the problem of uncalibrated cameras rectification,a direct and efficient three-step method of stereo rectification is proposed.To minimize the relative rotation between both cameras,the algorithm was decomposed into three-steps to limit the distortion.Considering the uncertainty of epipolar geometry measure introducing by the extimation of fundamental matrix,we propose an epipolar distance transform measurement.Results show that the new measure is more appropriate for image rectification.The three-step algorithm has obtained an accuracy comparable result both in estimation error and visual effect,especially when the initial epipolar lines are far from horizontal.2)In the local methods,the accuracy of support window influences the matching effect directly.Thereby we propose a ‘butterfly-shaped' adaptive window method to improve the accuracy of less texture and discontinuous regions,whose shapes change flexibly,and are sufficient to the complicated and various local images with low calculation complexity.Experimental results indicate that our proposal characterizes the image local structure information effectively.In the end,we take two steps: Union Jack shape voting and bilateral filtering algorithm as a post-processing step to alleviate the depth edge expansion problem.3)Utilizing the histogram distribution statistical characteristics,a new cost aggregation strategy based on histogram is proposed,where the space weights and color weights are calculated within the annular blocks.The histogram based matching cost keeps the accuracy of algorithm and improves the matching efficiency.But it leads to mismatches within the similar spatial statistical distribution areas.Like histograms,spatiograms are efficient to compute,and retain the global positions of the pixels.There is a one-to-one correspondence relationship with the image,and then the matching cost based on spatiogram improves the accuracy in the similar spatial statistical distribution areas.Experimental results demonstrate the efficiency and accuracy of our proposed method,meanwhile the histogram-based and the spatiogram-based matching cost aggregation methods are both robust to noise,and improve the matching accuracy within the textureless and discontinuous areas.4)Local methods typically assume that all pixels inside the support window have the same disparity,which leads to disparity misestimation for slanted surfaces.We propose an adaptive damping factors-based permeability filtering method,which is established on the adaptive ‘butterfly-shaped' support window.The adaptive damping factors are calculated adaptively,which relax the local smooth assumption and permit diaparity variation for similar points within slanted surfaces.Experimental results demonstrate that the proposal improves the mismatching of the slanted surfaces and even for the cases where the surfaces do not contain sufficient textural information.5)We propose a disparity candidate subset method to hold the efficiency.By the calculation of the coarse matching cost,we make up the disparity candidate subset by selecting the disparities in order,which satisfy the minimum coarse matching cost.Experimental results demonstrate that this method reduces the computation quantity fundamentally by reducing the search range of disparity,and enhances the accuracy of matching algorithm by preventing the unnecessary support weights to the aggregation.6)The image intensity-based stereo matching is sensible to noise and illumination variation,which cause ambiguous in less texture regions and discontinuous regions.An adaptive epipolar distance transform is proposed,which converts image intensity values to a relative location inside a planar segment along the epipolar line,thus we match with the image structure.The local segmentation scale factor and transformation parameter are calculated by the object's scale algorithm and the block discontinuous map.Adopting the images obtained by the proposed transform to estimate disparity,experimental results demonstrate the effectiveness in low-texture regions and discontinuous regions,such as edges and noises.