特征聚类与霍夫投票约束下影像匹配粗差剔除  

Outlier removal via feature clustering and Hough voting

作　　者：姜三 罗海涵 危迟 李清泉 江万寿[4] 王力哲 JIANG San;LUO Haihan;WEI Chi;LI Qingquan;JIANG Wanshou;WANG Lizhe(School of Computer Science,China University of Geosciences,Wuhan 430074,China;Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ),Shenzhen,Guangdong 518060,China;Shenzhen Smart Mapping Technology Co.,Ltd,Shenzhen,Guangdong 518052,China;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430072,China)

机构地区：[1]中国地质大学(武汉)计算机学院,武汉430074 [2]人工智能与数字经济广东省实验室(深圳),广东深圳518060 [3]深圳智绘科技有限公司,武汉430074 [4]武汉大学测绘遥感信息工程国家重点实验室,武汉430079

出　　处：《测绘科学》2023年第11期69-81,共13页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(42001413,42371442);人工智能与数字经济广东省实验室开放基金项目(GML-KF-22-08)

摘　　要：针对一些地物几何形态复杂的大比例尺无人机遥感影像,传统影像特征匹配算法难以高效处理粗差,影响后续定向和建模的问题,本文基于特征点聚类与霍夫投票策略,提出一种面向复杂场景的自动粗差剔除方法:①基于DBSCAN密度聚类将初始匹配点划分为几何关系一致的对应簇,避免单一几何变换模型无法模拟影像对复杂几何变形的问题;②使用RANSAC严格几何估计剔除各簇对的外点,并对簇中特征点构造基于三角形面积比值的霍夫投票,以此建立簇对之间的隐式几何变换模型,从而达到区分内点与外点的目的;③利用精匹配结果构造基于Delaunay三角网的匹配扩展,恢复误剔除的匹配点。利用近景数据集和无人机影像,并与其他粗差剔除算法进行了试验和对比分析。结果表明,本文所提出算法能够实现高外点率下的稳健粗差剔除,并很好地满足无人机影像SFM三维重建需求。Gross error removal has always been a core difficulty in image feature matching.For large-scale UAV remote sensing images with complex geometric shapes,traditional image feature matching algorithms are difficult to efficiently handle gross errors,which affects subsequent orientation and modeling.Based on feature point clustering and Hough voting strategy,this paper proposes an automatic and accurate outlier removal method for complex scenes.First,based on DBSCAN density clustering,the initial matching points are divided into corresponding clusters with consistent geometric relationships,which can avoid the problem that a single geometric transformation model cannot simulate the image’s complex geometric deformations.Second,the outliers of each cluster pair are eliminated by using the RANSAC-based geometric estimation,and the Hough voting based on triangle area ratio is constructed for the feature points in the clusters to establish an implicit geometric transformation model between the cluster pairs,which achieves the purpose of distinguishing outliers from inliers.Finally,the refined matching results are utilized to construct a matching extension module based on Delaunay triangulation to recover the mistakenly removed matches.Experiments and comparative analyses are conducted by using both close-range UAV datasets and with other outlier removal algorithms.The results demonstrate that the proposed algorithm can realize robust outlier removal even under high outlier rate and can meet the requirement in SFM-based 3D reconstruction for UAV images.

关 键 词：特征匹配 粗差剔除 特征聚类 霍夫投票 匹配扩展 运动恢复结构 

分 类 号：P231.2[天文地球—摄影测量与遥感]