Research On Dense Stereo Matching Algorithm Based On PatchMatch

Binocular stereo vision technology as the theoretical basis of robot autonomous navigation,virtual reality and augmented reality,medical image reconstruction,Autodriving and other applications,of which stereo matching algorithm is not only the most difficult step,but also the most important core.After lots of theoretical research and application practice,the accuracy of stereo matching algorithm has been improved significantly,however,the stereo matching algorithm still cannot solve the error matching problem at weak texture region of the image.To solve this problem,we analyzes the basic principle of stereo matching algorithm based on PatchMatch optimization idea,and compare the advantages and disadvantages of various algorithms that solving the mismatching problems in low texture region,we propose a dense stereo matching solution based on pixel classification and improved LocalExp(Local Expansion Move).The main work is as follows:(1)A local cross-window matching algorithm.In the initialization of pixel 3D disparity label of LocalExp algorithm,which has some unreliable label values by the initial process based on PatchMatch stereo matching algorithm,the constraint of disparity range of pixel points and visual constraint information under left and right views are introduced to generate more feasible 3D disparity labels.A cross-window model is designed to extract more local windows to calculate matching cost and label sharing between windows.Based on the model,a local window energy function which integrates multi-dimensional weights is defined,in order to ensure the stability and smoothness of the initial cost energy process of the matching cost,a constraint function is used to deal with the outliers instead of the traditional truncation.(2)Matching algorithm based on pixel information.Firstly,SNIC segmentation algorithm and Meanshift algorithm are used to segment the original image,and combine their information as the prior knowledge of the disparity label generation mechanism.Secondly,texture information of unstable pixel points in the disparity image are separated by texture splitter and left-right consistency splitter.With the use of pixels classifier,the optimization principle of PatchMatch and GraphCut algorithm,we redefine the 3D disparity proposal as the label generation scheme of proposed algorithm.(3)Reduce the error matching rates of weak texture region.A disparity refinement method based on cross line structure is proposed to improve the error disparity label in weak texture region.This method that uses pixels classifier to optimize error pixels with different classification information.Finally,the algorithm verification is carried out in the data set of Middlebury.The experiment results shows that proposed algorithm improves the matching effect in weak texture and non-weak region of image,while retaining the universality for the texture region.Compared with other methods on the Middlebury 2006 Stereo Vision Database,the algorithm in this paper achieved the lowest average error matching rates.