Research On Several Issues Of Oblique Stereo Image Matching

Oblique photogrammetry technology has been becoming one of the high-tech in international surveying and mapping field in recent years.The aerial oblique photography system captures nadir and oblique images by multiple sensors mounted on one flight platform introduced users a real and intuitive world that fits the human eye.It is one of the effective ways to obtain spatial data,and has unique advantages in the field of three-dimensional(3D)urban construction.Due to the significant change of the sensor's view angle during image acquisition,there are problems such as large geometric and radiometric distortion between stereo images and occlusion of corresponding objects etc.,which increase the difficulty of automatically determining the corresponding features.The unreliable result of feature matching directly leads to the processes of triangulation and the ground object extraction on the oblique images more difficult than that on the traditional aerial survey images.Therefore,study on the reliable automatic matching algorithm for oblique stereo images is of great significance for promoting the rapid development of oblique aerial photogrammetry.Based on image local features,this thesis focus on matching stereo images with complex distortion,integrating multiple features matching and stratght line matching for oblique images.The main research contents are provided as follows:(1)Commonly used features including Harris,Maximally Stable Extremal Regions(MSER),Scale-Invariant Feature Transform(SIFT),and Features from Accelerated Segment Test(FAST)and KAZE were analyzed;the performances of matching methods on the basis of diferent features for image pairs with geometric distortion(translation,scale and affine,etc.)and radiometric distortion(noise and brightness,etc.)were compared.(2)Focusing on matching for images with complex distortion,Weighted ?-SHape(W?SH)local feature was introduced into stereo image matching;Due to the disadvantages of small number of features and noise sensitive,2D Discrete Wavelet Transform(2D-DWT)was applied to improve W?SH-based matching method and three methods including WWF,IWWF and LIWWF were proposed.Experimental results shown that the W?SH-based matching methods achieved higher accuracy than MSER-based matching method in most cases;WWF,IWWF and LIWWF had better performance compared with W?SH in terms of both number of matches and matching accuracy;IWWF was the most stable method among MSER,W?SH,improved-W?SH;(3)Considering that the normalized cross-correlation(NCC)algorithm can not directly determinate correscondences on images with distortions such as rotation,scale and affine,Affine-Invariant NCC(AINCC)matching algorithm was proposed on the basis of affine transformation relationship between the neighborhood of feature points.Two applications of AINCC were tested on a series of analogous images.A matching method intergrating affine-and scale-invariant features was proposed for oblique images matching.The SIFT features in the neighbour area of MSER maches were matched by using AINCC algorithm,and the Neighbour Support Strenth(NSS)strategy was provided to remove error matches,then an iterative propagation matching based on local homography constraint was proposed to obtain a sufficient number of uniformly distributed matches.The optimal threshold value for the AINCC coefficient and the NSS coefficient in the initial match was discussed.The registration results for images with different geometric distortions shown that the proposed algorithm was better than the state-of-the-art matching algorithms and was a robust matching method that achieves high registration accuracy for oblique images.(4)In order to solve the problem that the existing line matching method quite relies on the image gray information,a line matching based on the spatial structure information composed of the distance between the lines and the midpoints,the rotated angle and the overlap degree was proposed.The proposed method uses the feature points to estimate the projective transformation model between images,and obtains line matches based on the distance between the lines and the midpoints of lines,then achives one-to-one matching by error matches removing and matches optimizing.Experimental results shown that the accuracy of line matches of proposed method was high,and the matching results were almost unaffected by the point matching result or image noise.