A Multi-View Image Matching Method Based On Semi-Global Optimization And Its Application

The images with high overlap covering the same area provide a large number of redundant information for image matching, thus multi-view image matching becomes the research hotspot in the field of digital photogrammetry. Multi-view images have high overlap, so how to fully use the multi-view images’redundant information in matching process, is the key to multi-view image matching. Aiming at reliable image matching, this dissertation proposes two kinds of multi-view image matching methods based on semi-global optimization, and presents a three dimensional reconstruction method using the generated dense photogrammetric point cloud. The main content of this paper includes:(1) Aiming at the image matching on difficult area such as surface discontinuities and occlusion, repetitive texture, textureless, this dissertation proposes a semi-global multi-view image matching method based on tracking corresponding points in image space. It uses semi-global matching to all of rectified stereopair in order to obtain the corresponding points, and then merges corresponding points between multi-view images by means of tracking, finally uses multiple forward intersection iterative optimization to realize the fusion of matching results in object space to generate an integration of matching result. Multi-view image matching has great matching redundancy, it can improve the reliability of image matching, and multiple forward intersection iterative optimization can improve the intersection accuracy, which provides reliable dense point cloud for three-dimensional reconstruction.(2) As the single stereo matching in the occluded area is easy to produce ambiguity match, and is easy to produce" multiple match" in repetitive texture and textureless area, this dissertation proposes a semi-global multi-view image matching method based on geometric constraint in object space. Based on the geometric relationship, the geometric constraint in the object space can guide the multi-view images to match at the same time, and the semi-global optimization can reduce wrong match. With a coarse-to-fine propagation strategy using image pyramid, matching result from high level of image pyramid are used to constrain the matching of low level of image pyramid to reduce the wrong matches caused by ambiguity matching, while reduce the memory consumption and the computational complexity.(3) Based on reliable and dense photogrammetric point cloud, a Possion surface reconstruction method is used to reconstruct the geometry topology structure of the scene, after that the corresponding images are mapped to the3D models as texture to get realistic scene model.This dissertation proposes two kinds of multi-view image matching methods based on semi-global optimization. The theory and algorithm of the methods are implemented using Visual Studio2010. Experimental results show that the proposed methods can provide reliable dense point cloud for three-dimensional reconstruction, which provide a portable way for high precision three-dimensional reconstruction.