倾斜摄影测量多元影像特征融合匹配策略与方法研究

Automatic and robust stereo matching has become a bottleneck problem limiting the practical applications of three-dimensioanl(3D)reconstruction in oblique photogrammetry. Furthermore, with the development of usage of digital surface model (DSM) in many fields, the precise and complete object surface are necessary for 3D reconstruction. Aiming at the aobve problems, this project will study and establish an integrative theories, methods and techniques used for stereo matching and 3D reconstruction of oblique photogrammetry. Among the integrative system, multiple features including points, straight or curved lines and regions will be selected as the matching primitives. The combinatory system includes the following aspects: (1) after analyzing and selecting affinely invariant regions with complementary features in some aspects,the project will explore a reliable sparse matching methods with higher correct-match number, correct-match rate and matching precision;(2) in this stage, the project will firstly study a similarity measure invairant to intensity changes and disparity discontinuity and projectively invariant matching method for straight or cuverd line features, then the SIFT feature points and straight or curved lines will be further selected and used to maching propagation in order to get matches as many as possible,simply because SIFT (Scale Invariant Feature Transformation) algorithm can detect feature points with dense distribution and higher positioning accuracy. On the other hand, straight or curved line features can represent the scene structure with higher level than feature points, which will contribute to subsequent 3D reconstruction. Based on the above methods, the scene-adaptive, multi-level propagating strategies by combining multiple primitives will be studied and established; (3) by integrating the theories and methods established in the previous steps systematically,the project will build the precise 3D reconstruction and evaluation system for practical applications by integrating multiple correspondences.

倾斜立体影像的匹配已成为制约数字摄影测量三维重建等实际应用的瓶颈问题。此外，随着数字表面模型（DSM）在诸多领域应用的不断发展，物体表面结构的精确性与完整性已成为三维重建的一个重要方面。针对上述问题，项目在特征提取、稀疏匹配及匹配传播的完整理论和技术框架下，通过融合点、线及区域特征，研究构建倾斜摄影测量多元影像特征融合匹配策略与方法。包括：（1）分析选取具有互补特性的仿射不变区域特征，研究具有较高匹配数量、匹配正确率及匹配定位精度的高可靠初始稀疏匹配方法；（2）在研究顾及亮度变形及视差不连续的相似性函数模型及线特征的投影不变匹配策略与方法的基础上，通过进一步融合具有密集、高精度检测特性的SIFT特征点及具有高层描述信息的线特征，研究构建融合多元影像特征的场景自适应的多层次、多约束密集匹配传播策略与方法；（3）对上述理论和方法进行系统集成，构建融合多元同名特征的三维重建应用及评价体系。

倾斜立体影像因其成像几何模型稳定、覆盖范围广以及纹理信息丰富，在城市真三维重建中发挥着关键作用。然而，由于传感器在获取影像过程中视角发生显著变化，导致影像间存在较大的几何和辐射畸变、同名区域遮挡等问题，加大了计算机自动立体量测同名像点匹配的难度。因此，研究此类影像的可靠匹配算法，无论是对提高城市真三维重建的效率，还是对推动数字摄影测量自动化进程，均具有重要意义。.针对上述问题，项目在特征提取、稀疏匹配及匹配传播的完整理论和技术框架下，通过融合点、线及区域特征，研究构建倾斜摄影测量多元影像特征融合匹配策略与方法。具体研究了（1）基于Harris-Affine仿射不变特征的特征匹配方法；研究多重几何条件的约束以及由粗到精的多级匹配策略，实现同时适用于平面及3D场景的高精度，高稳定性的密集宽基线影像匹配算法；（2）研究不同的仿射不变特征引入倾斜影像匹配过程并进行对比分析，研究融合仿射及尺度的倾斜影像匹配算法；（3）研究融合仿射及尺度的倾斜影像稠密匹配算法，实现大畸变影像高精度快速稠密匹配；（4）研究基于空间结构信息的直线匹配方法。.本项目在《Isprs Journal of Photogrammetry & Remote Sensing》、《IET Computer Vision》、《Journal of Electronic Imaging》等知名外国期刊上发表SCI索引论文4篇，在国内知名期刊发表论文8篇，其中EI检索论文3篇，申请并授权发明专利一项。培养博士2名，硕士4名。.本课题深入研究了倾斜摄影测量多元影像特征融合匹配策略与方法，构建了融合多元同名特征的三维重建应用及评价体系中的重要模块，为物体表面结构三维重建的精确性与完整性提供了基础和保证。