摘要
针对无人机多视角三维立体建模过程中,无序影像集(或其子集)具有随机性,进而影响建模稳健性的问题,该文提出一种基于交叉验证的影像选择方法,通过随机抽样和方差分析选择影像。首先,为确保最大程度地利用影像集中的信息,同时避免过度依赖特定的影像,对原始影像集进行随机采样分组。其次,利用影像几何特征,并联合影像特征匹配数量及特征分布情况来选择每组内最佳候选初始影像对。然后,再根据初始影像对选择场景重建所需后续影像。最后,对所有分组进行模型评定以及模型选择后利用全局优化方法对重建结果优化。对比实验结果表明,在3组小规模数据集上重投影误差分别为0.436、0.849、0.483像素,较Colmap方法分别降低45.6%、33.5%和26.2%。在大规模数据集上的重投影误差为0.338像素,精度优于对比方法,且本文方法对于实验场景具有更好的稳健性。
This paper proposes a cross validation based image selection method to address the issue of randomness in the unordered image set(or its subset)during the multi view 3D modeling process of UAV,which affects the robustness of the modeling.The method selects images through random sampling and variance analysis.First,to ensure maximum utilization of the information within the image set while avoiding over-reliance on specific images,the original image set is randomly sampled and divided into groups.Next,based on the geometric features of the images,the method selects the best candidate initial image pairs within each group by considering the number of feature matches and their distribution.Then,subsequent images needed for scene reconstruction are chosen based on the selected initial image pairs.Finally,after model evaluation and selection for all groups,a global optimization method is applied to refine the reconstruction results.Comparative experimental results show that,for three small-scale datasets,the re-projection errors are 0.436,0.849,and 0.483 pixels,respectively,which represent reductions of 45.6%,33.5%,and 26.2%compared to the Colmap method.On a large-scale dataset,the average re-projection error is 0.338 pixels,which outperforms the comparison methods,demonstrating better accuracy and robustness for the experimental scenarios.
作者
杨涛
李佳田
阿晓荟
李思旭
YANG Tao;LI Jiatian;A Xiaohui;LI Sixu(Faculty of Land Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China)
出处
《测绘科学》
北大核心
2025年第6期123-133,共11页
Science of Surveying and Mapping
基金
国家自然科学基金项目(41561082)。
关键词
三维重建
影像选择
特征分布
交叉验证
增量重建
3D reconstruction
image selection
feature distribution
cross-validation
incremental reconstruction
