Research On The Image Matching Methods In 3D Modeling Of Oblique Images

Image matching is the core of aerial photogrammetry.The quality of image matching has great influence on the accuracy of the final 3D model.Sparse matching affects the accuracy of aero triangulation,while dense matching affects the accuracy of 3D point clouds models.Compared with the traditional photograph,the oblique photograph has more information and can get better 3D model,but it also brings new problems.In this paper,we research on the matching methods for oblique images in order to deal with the difficulties in matching of oblique images.The main research contents in this dissertation are listed as follows:(2)For sparse matching,the traditional sparse matching strategy is first introduced.Two matching strategies are introduced in oblique photography:using an affine invariance matching algorithm and correcting the affine deformation by correcting the image before matching.The two matching strategies are simulated and compared with the efficiency and the number of corresponding points.(3)For the sparse matching by correcting the image before matching,a method with second correction is proposed to improve the performance.Through simulation,it is confirmed that the new method can get more matching points.(3)For dense matching,some traditional matching methods are introduced.A triangulation constraint matching algorithm which uses matching points obtained by sparse matching as prior knowledge is proposed.The corner feature points are introduced to divide the triangles which contain disparity discontinuity.(4)Aiming at the common areas of oblique photography,the right angle corners of buildings commonly used in cities,factories and other scenes is selected as the feature points to be detected.Harris algorithm is used to extract feature points,and canny algorithm and Hough transform algorithm are used to extract straight lines.The location and direction of feature points are determined by information of points and straight lines.The SIFT descriptor is modified according to the location and direction of the feature points.Simulation results show that this method can effectively reduce the number of triangles that are difficult to match.