多角度遥感影像建筑物区域立体匹配及重建研究

发布时间：2018-06-29 22:01

本文选题：多角度遥感影像 + 有理函数模型优化　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：随着卫星立体成像能力的逐渐成熟,卫星应用数据需求也逐渐受到人们的广泛关注,卫星数据也从传统的立体像对过渡为多角度影像。与传统的二维平面图像不同,通过卫星影像和相应的参数文件,可以对影像覆盖区域完成三维重建。建筑物作为人类生活区域中最常见的一类地物,一直与我们息息相关。因此,本文利用多角度遥感影像,对建筑物区域完成立体匹配及重建,主要围绕遥感影像通用成像模型,立体匹配及重建这三部分展开,其中,对通用传感器模型的讨论主要侧重于有理函数模型优化,立体匹配则针对建筑物自身特点采用针对性的技术手段完成影像联合立体匹配工作,在完成上述流程的基础上,利用基于有理函数模型的最小二乘算法和DSM数据融合插值技术,实现对建筑物区域立体重建。在整个立体重建系统流程中,遥感影像传感器模型建立了立体影像二维图像坐标到三维大地坐标的坐标变换关系。本文从最常见的广义传感器模型——有理函数模型入手,分析其优势与不足,介绍了两种经典模型优化算法。进一步的,针对优化过程中控制点不足及非均匀分布问题,提出解决方案;最后,结合经典算法,采用虚拟地面控制点,实现了有理多项式系数的精化,为后文近似核线影像生成和立体重建奠定了基础。在研究立体匹配算法之前,首先对多角度影像添加核线约束,获得近似水平核线影像,保证多角度影像中同名像点处于行搜索范围内。进而,以中间影像为基准影像,针对建筑物区域包含大量的角点及边缘信息,因此,采用点特征提取与匹配技术确定目标影像匹配搜索范围。同时,对直线特征提取过程做出了针对性的改善,利用线特征匹配视差及线段分布位置关系,获得最大最小视差分布图,并以此约束区域匹配过程,建立基准影像与左右目标影像间稠密视差图的生成。最后,为了实现真实场景下建筑物区域立体重建,首先研究了基于有理函数模型的最小二乘算法,实现匹配视差图到三维信息的转换,获得稠密DSM点云数据。然后针对匹配过程中匹配不完全导致DSM数据值部分缺失问题,在物方空间实现DSM数据融合,并在保持建筑物边缘特征的基础上完成插值等后处理过程。与同一场景下Lidar数据以及立体像对数据相比较,获得较好的实验结果。
[Abstract]:With the maturity of satellite stereo imaging ability, the demand for satellite application data has been paid more and more attention, and satellite data has been transformed from traditional stereo image to multi-angle image. Different from the traditional two-dimensional plane image, the 3D reconstruction of the image covering area can be completed by satellite image and the corresponding parameter file. As one of the most common features in human living areas, buildings are closely related to us all the time. Therefore, this paper uses multi-angle remote sensing image to complete the stereo matching and reconstruction of the building area, mainly focusing on the general imaging model, stereo matching and reconstruction of remote sensing image. The discussion of the universal sensor model mainly focuses on the rational function model optimization, and the stereo matching is based on the above process, and the corresponding technology is adopted to complete the image joint stereo matching according to the characteristics of the building itself. The least square algorithm based on rational function model and DSM data fusion interpolation technique are used to realize the stereoscopic reconstruction of building area. In the whole system of stereo reconstruction, the model of remote sensing image sensor establishes the coordinate transformation relationship between 2D image coordinate and 3D geodetic coordinate of stereo image. This paper begins with the most common generalized sensor model-rational function model, analyzes its advantages and disadvantages, and introduces two classical model optimization algorithms. Furthermore, to solve the problem of deficiency and non-uniform distribution of control points in the optimization process, a solution is proposed. Finally, the rational polynomial coefficients are refined by using virtual ground control points in combination with classical algorithms. It lays a foundation for the generation of approximate kernel line image and stereo reconstruction. Before the stereo matching algorithm is studied, the kernel line constraint is added to the multi-angle image firstly, and the approximate horizontal kernel line image is obtained to ensure that the image with the same name in the multi-angle image is within the range of row search. Furthermore, with the intermediate image as the reference image, the building area contains a lot of corner and edge information, so the point feature extraction and matching technology is used to determine the target image matching search range. At the same time, the line feature extraction process has been improved, using line feature matching parallax and line segment distribution relationship, the maximum and minimum parallax distribution map is obtained, and the matching process of constrained region is obtained. Build the dense parallax image between the reference image and the left and right target image. Finally, in order to realize the stereo reconstruction of the building area in the real scene, the least square algorithm based on rational function model is studied firstly, and the matching parallax map is transformed into 3D information, and the dense DSM point cloud data is obtained. Then to solve the problem of partial missing of DSM data value caused by incomplete matching, the DSM data fusion is realized in the physical space, and the post-processing process such as interpolation is completed on the basis of preserving the building edge features. Compared with Lidar data and stereo pair data in the same scene, good experimental results are obtained.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】