面向大规模三维重建的快速鲁棒集束调整算法  

Large Scale 3D Reconstruction-Oriented Fast and Robust Bundle Adjustment Algorithm

作　　者：曹明伟[1] 李书杰[1] 李尚林[1] 贾伟[1] 刘晓平[1] Cao Mingwei;Li Shujie;Li Shanglin;Jia Wei;Liu Xiaoping(School of Computer and Information, Hefei University of Technology, Hefei 230009)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机辅助设计与图形学学报》2017年第7期1303-1313,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61305093;61370167);国家科技支撑计划子课题(2012BAJ08B01)

摘　　要：集束调整是运动推断结构的核心,针对现有算法在大规模场景下易受外点影响,空间占用率过高和效率较低问题,提出一种快速鲁棒的集束调整(fast and robust bundle adjustment,FRBA)算法.首先,为了避免外点(outliers)的影响,采用Cauchy损失降低外点的权重,提高算法精度.其次,充分利用运动推断结构中三维点与摄像机之间的稀疏性对大规模集束调整进行稀疏分解,降低内存空间的使用.最后,根据稀疏分解后矩阵的固有特性,采用快速矩阵分解法求解正态方程的解.在合成数据集、BAL数据集和真实图像数据集上对FRBA算法进行测试,并与现有经典算法进行比较.实验结果表明无论在时间效率还是精度上,FRBA算法均处于领先位置.Bundle Adjustment is an important component in structure from motion,in order to overcome thedrawbacks existing in the robustness,accuracy and space time efficiency of the existing algorithms,a fast androbust bundle adjustment(FRBA)algorithm has been proposed.Firstly,in order to avoid the effects of outliers,we use the Cauchy loss function to potentially down-weight outliers for improving accuracy of the proposed method.Secondly,a large-scale bundle adjustment problem can be divided into some small ones by making use ofthe sparsity between3D points and the cameras for reducing the requirements of memory.Finally,according tothe inherent property of the matrix after its sparse decomposition,we use a fast matrix factorization algorithm tosolve the normal equation.The proposed FRBA algorithm was compared with some state-of-the-art methods onthe synthetic dataset,BAL dataset and real image datasets respectively.Experimental results show that the proposedFRBA algorithm outperforms the state-of-the-arts on both time efficiency,space efficiency as well as precision.

关 键 词：集束调整 三维重建 运动推断结构 摄像机跟踪 矩阵分解 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]