利用网络图进行高分辨率航空多视影像密集匹配

发布时间：2018-06-14 18:42

本文选题：高分辨率 + 多视影像　；参考：《测绘学报》2016年10期

【摘要】：提出了一种基于网络图的高分辨率航空多视影像密集匹配算法。首先利用影像间的重叠关系和方向确定候选立体像对并构建网络图;在立体像对密集匹配阶段,引入导向中值滤波采用由粗到精的改进半全局匹配(SGM)算法进行双向视差图生成;最后基于所有立体像对构建的网络图完成多视影像间的密集点云生成及融合。试验选取了ISPRS的Vaihingen航空影像和ISPRS/EuroSDR项目的苏黎世倾斜下视影像进行试验,结果表明:本文算法对高分辨率多视影像密集匹配是有效可行的,无论在匹配完整性、效率、精度上都能获取较好的结果,重建的密集点云平均反投影误差的中误差可以达到亚像素级精度,实际精度可以达到1.5倍GSD,并且在建筑物、植被、水体等视差不连续、弱纹理或重复纹理区域也取得了较好的匹配结果。
[Abstract]:This paper presents a dense matching algorithm for high resolution aerial multi-view images based on network graph. Firstly, the candidate stereo pairs are determined by the overlap relation and direction between images and the network map is constructed. In the phase of stereo image matching, the steered median filter is introduced to generate the bidirectional parallax map with the improved semi-global matching algorithm from coarse to fine. Finally, the dense point cloud generation and fusion between multi-view images are completed based on the network diagram constructed by all stereo pairs. The experiment selected ISPRS's Vaihingen aerial image and ISPRS / EuroSDR project's oblique downward viewing image in Zurich. The results show that the algorithm is effective and feasible for high-resolution multi-view image dense matching, regardless of matching integrity and efficiency. The average error of the reconstructed dense point cloud can reach sub-pixel level accuracy, the actual precision can reach 1.5 times GSDs, and the parallax is not continuous in buildings, vegetation and water bodies. The weak or repeated texture regions also have better matching results.
【作者单位】：武汉大学测绘学院;
【基金】：测绘地理信息公益性行业科研专项(201512008) 中央高校基本科研业务费专项资金(2015214020201)~~
【分类号】：P23

【相似文献】