基于多视立体视觉与Leiden图聚类的多视图直线匹配  

Multi-view Line Matching Based on Multi-view Stereo Vision and Leiden Graph Clustering

作　　者：兰泽清 王竞雪[1,2] 王丽芹[3] LAN Zeqing;WANG Jingxue;WANG Liqin(School of Geomatics,Liaoning Technical University,Liaoning,Fuxin 123000,China;Collaborative Innovation Institute of Geospatial Information Service,Liaoning Technical University,Fuxin 123000,China;Land Resource Surveying and Planning Institute of Liaoyang City,Liaoyang 111000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,阜新123000 [2]辽宁工程技术大学地理空间信息服务协同创新研究院,阜新123000 [3]辽阳市国土资源勘查规划院,辽阳111000

出　　处：《地球信息科学学报》2024年第7期1629-1645,共17页Journal of Geo-information Science

基　　金：国家自然科学基金面上项目(41871379);辽宁省兴辽英才计划项目(XLYC2007026);辽宁省应用基础研究计划项目(2022JH2/101300273)。

摘　　要：直线特征的精确匹配对三维模型的重建与优化具有重要意义。但是传统的双视图直线匹配由于视图数量较少导致直线匹配结果鲁棒性较差,且对于存在断裂的直线提取结果,同一直线在不同影像上提取直线的数目不等,导致直线匹配结果完整性较差。针对上述问题,本文提出多视图立体视觉(Multiple View Stereo,MVS)与Leiden图聚类相结合的多视图直线匹配算法。该算法首先使用直线提取算法和MVS三维重建算法对输入的多视影像分别进行直线信息提取和多视图三维信息提取,得到各视图上的直线与覆盖多视图场景的稠密三维点及其相关信息,并以此构建直线描述符;接着通过三维直线投影角度约束、点-线位置关系约束及同名点约束对匹配候选进行筛选,并根据各视图上直线间的相似性关系以各视图上的直线为节点,直线间的相似性得分为边权重构建无向图,同时剔除无向图中由单个节点构成的连通分量,得到连通分量集合,即初始匹配结果;最后对各个连通分量的节点基于同视图共线约束进行重构,得到子无向图,并对子无向图节点使用Leiden算法进行聚类。最终在多视影像上得到精确的直线匹配结果。本文选取大疆无人机拍摄的2组影像以及网上公开的1组影像进行实验。实验结果表明,采用本文算法获得的直线匹配结果较对比算法在直线匹配数目以及匹配正确率上均有所提升。Accurate matching of line features is of paramount importance in the reconstruction and optimization of three-dimensional models.However,traditional dual-view line matching encounters challenges due to a limited number of views,resulting in suboptimal robustness in line matching.For line extraction results with breaks,the number of lines extracted for the same line on different images is different,resulting in poor integrity of straight line matching results.To address these issues,this paper proposes a multi-view line matching algorithm that combines Multiple-View Stereo(MVS)and Leiden graph clustering.The algorithm commences by employing the line extraction algorithm and the MVS three-dimensional reconstruction algorithm on input multi-view images for line information extraction and multi-view three-dimensional information extraction,respectively.This process yields lines on each view,dense three-dimensional points encapsulating the image scene,and the correspondence between object-side three-dimensional points and their corresponding image-side two-dimensional points.Building upon this foundation,the algorithm constructs line descriptors in the image domain by considering lines and their matching point sets within their neighborhoods.Subsequently,leveraging the three-dimensional line projection angle constraints,point-line position relationship constraints,and corresponding point constraints,the algorithm filters matching candidates based on these three geometric constraints.Harnessing the similarity relationships between lines on each view,an undirected graph is constructed.Here,lines on each view serve as nodes,and the similarity scores between lines act as edge weights.Simultaneously,connected components composed of single nodes are removed from the undirected graph,resulting in the set of connected components that represent the initial matching results.In the final stage of this process,nodes of each connected component are reconnected based on same-view collinear constraints,forming many sub-undirected gr

关 键 词：多视图直线匹配 特征匹配 几何约束 无向图 稠密三维点 多视图立体视觉 Leiden算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]