【作者】 张鑫；

【导师】 王竞雪；

【作者基本信息】 辽宁工程技术大学 ， 测绘工程（专业学位）， 2022， 硕士

【摘要】 近年来,高分辨率遥感影像在国民经济建设过程中的应用越来越广泛,其中基于遥感影像的三维重建技术扮演着重要角色,受到各个国家的密切关注,而获取可靠的三维信息是实现高精度三维重建的基础。本文旨在通过构建约束条件以实现匹配点的粗差剔除,从而获取高精度的密集匹配结果,并在高精度的密集匹配结果上实现三维信息提取的研究。针对现有由稀到密的加密匹配算法中,初始匹配点可靠性低将导致迭代匹配拓展过程存在较多误匹配,无法获取可靠的三维点信息,提出一种基于局部纹理和局部几何特征约束的密集匹配及三维信息提取算法。首先,利用尺度不变特征变换(Scale-invariant Feature Transform,SIFT)获取的匹配点约束直线匹配获得的同名直线构建虚拟匹配点集,结合虚拟匹配点集和SIFT匹配点集建立初始匹配点集;然后,利用指纹信息和梯度信息构建匹配点局部区域约束剔除较为明显的误匹配点,并利用匹配三角网构建局部几何约束剔除由相似纹理产生的误匹配点,得到优化后的可靠匹配点;接着,基于可靠匹配点构建的Delaunay三角网,以三角形重心为加密匹配基元,结合核线约束和仿射变换对其进行迭代匹配拓展,得到最终密集匹配点集;最后,以密集匹配点集为基础,采用联合投影射线法获取其三维点信息。选取4组资源三号卫星前视数据和后视数据进行实验,在4组数据上其平均匹配精度为91%,具有较好的匹配稳定性,并利用两组影像为代表进行解算高程和真实高程误差对比,结果表明,平均高程点误差为2.37 m,相对于对比算法具有更低损失,且其三维信息提取结果表现更加连续和平滑。上述研究表明,局部特征约束充分考虑到了匹配点间的几何相似性和纹理相关性,接着针对匹配点之间的全局关系,即匹配点之间的空间相关性,提出一种基于均匀超图结构一致性约束的密集匹配及三维信息提取算法。首先,依据同名直线和SIFT匹配点构建初始匹配点,并利用局部影像信息约束剔除部分误匹配点得到初始可靠匹配点;其次,以初始可靠匹配点为基础,依据参考影像匹配点索引,在两影像上构建3-均匀超图结构,分别遍历计算其超边权重,并利用高斯核函数将对应超边权重结合,以构建关联超图;然后,以关联超图为基础,利用高阶主聚类分析算法迭代计算匹配点评分以获取最优可靠匹配点;最后,以可靠匹配点为基础获取密集匹配点集合,在此基础上实现三维点信息提取。选取4组资源三号卫星前视数据和后视数据进行实验,在4组数据上其平均匹配精度为92%,并利用两组影像为代表进行解算高程和真实高程误差对比,结果表明,平均高程点误差为2.30 m,相对于对比算法,其在高程精度上和高程平滑性上更具有优势。该论文有图31幅,表12个,参考文献61篇。更多还原

【Abstract】 In recent years,high-resolution remote sensing images have been applied more and more widely in the process of national economic construction.Among them,3D reconstruction technology based on remote sensing images plays an important role and is paid close attention to by various countries.Obtaining reliable 3D information is the basis of realizing high-precision3 D reconstruction.The purpose of this paper is to obtain high-precision dense matching results by constructing constraint conditions to eliminate the gross error of matching points,and to realize the research of 3d information extraction based on high-precision dense matching results.To avoid the problem of mismatches caused by initial matched points that may contain false matches during iterative dense matching based on corresponding points,a dense matching and3 D information extraction algorithm based on local texture and local geometric feature constraints is presented.Firstly,to increase the number of initial matching points and expand the coverage range of initial matching points,the initial set of matched points containing the matched SIFT(Scale-invariant Feature Transform)points and virtual corresponding points is constructed,where the virtual corresponding points are generated from the intersections of corresponding lines obtained by the line matching algorithm based on the matched SIFT points constraint,Secondly,initial matched points set is checked for removing the false matches using local image information and local geometry constraints in turn.This process first uses the local texture feature constraint constructed based on fingerprint information and gradient information to eliminate the mismatched points with low similarity,and then uses the local geometric constraint constructed by Delaunay triangulation to remove the false matches generated by similar textures,thereby obtaining the optimized set of reliable matched points.Finally,the Delaunay triangulation is constructed using reliable matched points,and the center of gravity of the triangles satisfying the area threshold is used as the matching primitive during the dense matching process,and matching based on the epipolar constraint and affine transformation constraint is performed iteratively to obtain the dense matching results.Finally,based on the dense matching point set,the joint projection ray method is used to obtain its 3D point information.This paper uses four sets of forward and backward viewing data of ZY-3 to perform experiment,in the four groups of data on the average matching accuracy is 91%,has good matching stability,and by using two sets of images represented by calculating elevation and real elevation error comparison,the results show that the average elevation point error is 2.37 m,compared with the contrast algorithm has lower losses,The results of 3D information extraction are more continuous and smooth.The above research shows that the geometric similarity and texture correlation between matching points are fully taken into account by the local feature constraint.Then,aiming at the global relation between matching points,namely the spatial correlation between matching points,a dense matching and 3D information extraction algorithm based on uniform hypergraph structure consistency constraint is proposed.Firstly,the initial matching points were constructed based on the corresponding lines and SIFT matching points,and the initial reliable matching points were obtained by eliminating partial false matching points with local image information constraint.Secondly,based on the initial reliable matching points and the reference image matching point index,the 3-uniform hypergraph structure was constructed on the two images,and the hypergraph was constructed by ergodic calculation of the hyperedge weights,and the corresponding hyperedge weights were combined with the Gaussian kernel function.Then,based on the associated hypergraph,the high order main cluster analysis algorithm was used to calculate the score of the matching points iteratively to obtain the optimal reliable matching points.Finally,based on reliable matching points,a dense set of matching points is obtained,and3 D point information is extracted.This paper uses four sets of forward and backward viewing data of ZY-3 to perform experiment,The average matching accuracy of the algorithm is 92% on four sets of data,and the error comparison between the calculated elevation and the real elevation is carried out by using two sets of images as representatives.The results show that the average elevation point error is 2.30 m,which has more advantages in elevation accuracy and elevation smoothness compared with the comparison algorithm.更多还原