Research On Fast Global Stereo Matching Algorithm

Binocular stereo vision is one of the important fields of computer vision research,which was widely used in robot navigation,unmanned vehicles,3d reconstruction and 3d scanning,target tracking,and many other fields.Stereo matching is the core of the binocular stereo vision system and its efficiency and accuracy directly affects the efficiency and accuracy of binocular stereo vision system.With the development of binocular stereo vision system,demand of real-time and high precision of binocular stereo vision system is very urgent.However,binocular stereo vision systems with hardware-accelerated can be improved to a certain extent and many high precision algorithms with hardware-accelerated still can't achieve the real-time request.So,there's much significance on the study of rapid high-precision stereo matching algorithms.At present,the stereo matching algorithms mainly include the following four steps:matching cost calculation,cost aggregation,disparity optimization,disparity refinement and are mainly divided into local algorithms and global algorithms depending on whether there's a disparity optimization.Local algorithms have high speed because of having no disparity optimization,but its accuracy is low.Some improved algorithms still can't satisfy the accuracy requirement of high precision fields and even have more time consumption than global algorithms.Global algorithms based markov random field have high accuracy because of having disparity optimization,but its speed is low.Recently,adaptive random walk with restart was introduced into the disparity optimization.It can effectively solve the problem which the speed of global matching algorithm based on markov random field is slow.Census transform can transform the original image into binary image which is suitable for computer processing and can speed up calculation.What' more,Census transform can deal with false matching problem caused by photometric distortions and noise.However,Census transform also can cause mistake matching of similar texture.So,this paper proposed an improved Census transformation--Census transform based Equicrural Triangle,WEJBF.So,this thesis combines the adaptive random walk and the improved ETCensus transform to accelerate disparity calculation.Disparity refinement is the final step in stereo matching and is independent of the first three steps.Median filtering is the commonly used disparity refinement technology,but it can't correct the parallax figure and spreads the unreliable point of discontinuity to the surrounding pixels.For this problem,this thesis proposed a Joint Bilateral Filter Based on Wavelet Multi-scale Edge Detection,WEJBF.Finally,combining with the improved content,this thesis proposed a rapid global stereo matching algorithm based on ETCensus and WEJBF with high speed and accuracy.The main achievements and research contents of this thesis are as follows:(1)Research on binocular stereo vision system and stereo matching technologyThis part is the theoretical basis of the binocular stereo vision,mainly analyzing the camera imaging principle and the principle of binocular stereo vision and roughly analyzing the binocular stereo vision system modules.In addition,this thesis study the whole stereo matching principle and steps in more detail,analyzing and comparing the current popular stereo matching algorithm,summing up the advantages,disadvantages and problems at the same time.(2)Stereo matching algorithm based on ETCensus and WEJBFOn the basis of using adaptive random walk with restart and census transform to accelerate the optimization process,this thesis uses ETCensus transform and WEJBF(Joint Bilateral Filter Based on Wavelet Multi-scale Edge Detection)to solve the problem which census transform cause mistake matching of similar texture and the traditional median filtering can't correct the parallax figure and spreads the unreliable point of discontinuity to the surrounding pixels.(3)Validation of the test platformThrough the comparison of experimental data with the algorithms of current internationally accepted Middlebury and KITTI test platform,verifying the effectiveness of the proposed algorithms.It can better balance matching speed and matching accuracy and can solve the problems of similar texture and the correction of discontinuities,especially the more disturbing images of outdoor scenes,being suited to binocular stereo vision system in real time.