摘要
针对当前大规模影像运动恢复结构中并行式光束法平差收敛性差、精度低的问题,该文提出了一种融合运动平均优化的并行式光束法平差方法。该方法将大规模影像间的匹配关系表示为共视图,利用图分算法将整个区块分为多个子区块,并对各子区块进行快速局部光束法平差;图分后,将各子区块作为顶点构建子区块聚类图;在聚类图上利用运动平均优化,获得各子区块的全局旋转和全局平移参数,同时保持子区块内部的刚体结构和不同子区块间的几何相关性;最后,融合各子区块的重叠区域得到完整场景,实现对整个区块的平差优化。使用大规模无人机影像和地面近景影像数据进行实验,结果表明,相比传统方法,该文方法计算效率提高了10倍;相比当前的并行式方法,该文方法在精度、收敛性和计算时间方面都有明显提高。
To solve the problem of poor convergence and accuracy of parallel bundle adjustment in large-scale image-based Structure from Motion,this paper proposes a parallel bundle adjustment method using motion-averaging optimization.Matching relationships between large scale image sets are denoted as a view graph which will be cut into several sub-block subsequently.Local bundle adjustment is executed on every sub-block in a much shorter time.A new cluster graph whose nodes are sub-blocks is established.Motion-averaging optimization is utilized on cluster graph to optimize global rotations and translations of sub-blocks,which helps hold rigid structure inside every sub-block and geometric relationships between sub-blocks.Global bundle adjustment for the whole scene is completed after fusion of sub-blocks’overlapping areas.Experiments are conducted on large-scale UAV image datasets and terrestrial datasets.Compared with traditional method,proposed method significantly improves computational efficiency by one order of magnitude.The proposed method also has advantages in accuracy,convergence and efficiency over former parallel method.
作者
吴凌辉
肖腾
李鸿辉
高旺
邓非
WU Linghui;XIAO Teng;LI Honghui;GAO Wang;DENG Fei(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China;School of Computer Science,HubeiUniversity of Technology,Wuhan 430068,China;Wuhan Tianjihang Information Technology Co.,Ltd.,Wuhan 430074,China;Science and Technology on Complex System and Control and Intelligence Agent Cooperation Laboratory,Beijing 100191,China)
出处
《测绘科学》
CSCD
北大核心
2023年第7期163-172,共10页
Science of Surveying and Mapping
基金
创新研究专项基金项目(2022C61540)
湖北省重点研发项目(2022BAA035)。
关键词
并行式光束法平差
运动平均
子区块聚类图
运动恢复结构
parallel bundle adjustment
motion-averaging
sub-block cluster graph
structure from motion
