Research On Image Matching Between Oblique Images Baesd On Third Order Moment And Propagating Of Self-Adaptive Patch

Aerial high-altitude photography with tall and fixed space posture (vertical or small-angle photography) can only get low-resolution texture, and top part of the building, so it is difficult to meet data acquisition requirements of three-dimensional model. With advantages of flexible, fast, high resolution, the combination of the newly developed low-altitude UAV with multi-sensor integration technology, can meet the needs of the current three-dimensional model of city buildings. However, from the results of the market and research institutes, the efficiency of3D landscape model production is low and the model is rough, it is difficult to meet the actual needs of users. On the low altitude multi-sensor image, a point on object is shot simultaneously in many images, so this is a contradiction of "data are many as well as not enough". On one hand, the overlay rate of multi-angle images and redundancy are high; On the other hand, the actual utilization rate of images is low and three-dimensional model is rough, thus unable to meet users' needs. In this dissertation, starting from the actual production, a technical route is designed to extract affine-invariant features to guide three-dimensional model production of the landscape based on dense matching.The main points of the dissertation are as follows:Firstly, the basic principle of the inclined photography is explained, and the dissertation analyzes the difficulties in stereo matching of oblique image. Then, the article summarizes the existing matching model and state-of-art feature detection and description algorithms. Finally, the article summarizes the generic constraint methods.The article summarizes the geometric characteristics of projective transformation, affine transformation, similarity transformation under the group of projective transformation. Then, the article deduces the conditions of affine transformation to measure transformation, and reviews the usual affine invariants. For the reason of large geometric distortion, the disparity fracture, elevation is not smooth in the low altitude image of urban area, this dissertation proposes a affine invariant feature extraction algotithm based on third-order moment to overcome geometric deformation of image. The experiments show that the algorithm achieves good results, compared to state-of-art algorithms in3different groups of images.For the characteristics of large overlap and redundancy of oblique images, this article firstly analyzes the necessity and feasibility of filtering images in the match process. Then, the article proposes a multi-channel binary evaluation model in matching process for filtering images in real time to find the best match and optimize the performance of multi-chip matching model, in aim to find the "fewer but better" images to complete the matching process.For the presence of parallax disparity in images of buildings in urban areas, this article presents a propagating of self-adaptive patch dense matching algorithm guided by invariant feature. The algorithm adjusts the size of the surface element according to the characteristics of the parallax range within the window. When the surface is smooth, the patch gets larger. When the surface is uneven, the patch gets smaller, thus generating accurate, reliable3D point cloud model.The article summarizes the algorithms of Delaunay Triangulation Algorithm under large amounts of data. Based on insufficient run memory commonly due to the large amount of data triangulation, low utilization of CPU core, the paper presents a parallel algorithm based on multi-core platforms. The algorithm combines the OpenMP multi-thread programming techniques with parallel streaming data triangulation and releases run memory in real time, achieving good results.The experiments evaluate the method proposed by this article from visual testing and quantitative evaluation two expects based on standard aerial images and1plus4oblique images.