The Research Of Stereo Matching Algorithm Based On Feature Points

In the field of computer vision,stereo matching is one of the important research directions,it is crucial for 3D reconstruction.As we know,a good matching method should meet the requirements of high speed and high precision in order to guarantee the performance of practical and real-time.The whole process of stereo matching is based on the left and right views of real scene,so stereo matching is largely dependent on the left view image and right view image.The uncertainty of the left and right views will have a great impact on the performance of stereo matching,so the problem of stereo matching is a difficult problem to be solved.In recent years,stereo matching technology has been continuously under development and improvement,which is mainly divided into stereo matching based on feature points and that based on region.In this thesis,the main work of stereo matching based on feature points is as follows:(1).Aiming at the problem of illumination invariance in stereo matching,a stereo matching algorithm based on global color transfer and PCA-SIFT is proposed in this dissertation.Firstly,the new algorithm implements the global color transfer between left view and right view which have different illumination under the same scene,resulting in reducing the error caused by color difference.Secondly,the PCA-SIFT algorithm is used to extract the features information and match the feature points of the left view and the processed right view in order to matching left view and right view,and delete false matching points.Lastly,region growing algorithm is used to generate disparity map to complete the stereo matching.The simulation results show that the new algorithm has a good performance in stereo matching when the illumination changes.(2).Aiming at the problem of noise in stereo matching,this dissertation proposes a stereo matching algorithm based on feature points classification and improved ORB.Firstly,the feature points of left view and right view are detected based on the feature points detection of ORB algorithm,and a two-dimensional feature vector is formed based on the difference of pixels in feature point neighborhood and the final vector is obtained by combining the two-dimensional vector with the SURF descriptor.Secondly,the feature points are grouped and matched according to the two-dimensional vector,exact matched according to SURF descriptor,and delete false matching points.Lastly,region growing algorithm is used to generate disparity map to complete the stereo matching.The simulation results show that the new algorithm has a good performance in stereo matching under the noisy condition.(3).In view of the matching problem in the area of inclined plane,curved surface and weak texture,this dissertation presents a stereo matching algorithm based on wavelet transform and GLOH.Firstly,the new algorithm implements wavelet transform between left view and right view under the same scene in order to generate the low frequency and high frequency sub images of the left and right views,and the SIFT algorithm is used to extract the features,GLOH algorithm is used to generate feature descriptors.Secondly,rough matching and precise matching of GLOH feature descriptor are performed,and the false matching points are deleted.Lastly,region growing algorithm is used to generate disparity map to complete the stereo matching.The simulation results show that the new algorithm has a good performance when matching in the inclined surface,the surface and the weak texture region.