基于三阶矩和自适应物方面元传播的倾斜影像匹配研究

发布时间：2018-04-29 04:15

本文选题：倾斜影像 + 特征提取　；参考：《武汉大学》2014年博士论文

【摘要】：中高空航空摄影由于航飞高度大和固定的空间拍摄姿态(垂直或者小角度摄影)的原因,仅仅能获取建筑物的低分辨率纹理和顶部纹理信息,难以满足城市建筑物三维模型对纹理的要求。而低空无人机结合新近发展的多传感器集成技术,具有灵活、快速、分辨率高的优点,将其用来采集城市建筑物三维模型信息,可以满足城市建筑物三维模型构建的需求。但是,从目前市场和科研单位的成果上来看,三维景观模型生产效率低且粗糙,难以满足实际的用户需求。在低空多传感器影像上,一个物方点可同时成像于多张影像上,因此,这是一个数据“又多又少”的矛盾。一方面,多角度影像重叠度高,冗余信息大；另一方面,影像实际利用率低,生产的三维模型粗糙,无法满足用户需求。本论文从实际生产出发,设计出一套利用低空倾斜影像特征信息提取,引导密集匹配生成景观三维模型的技术路线。论文完成的研究内容如下：对倾斜摄影技术的基本原理进行说明,分析了倾斜影像立体匹配存在的难点；然后,对现有匹配模型进行系统分析、总结和分类；总结了现有的基于像方的特征角点检测算子、描述子和基于物方的立体匹配算法；对通用的匹配技术和约束方法进行归纳和总结。对射影变换群下的射影变换、仿射变换、相似变换几何特点进行系统总结；总结了各个变换的不变量和变换之间的自由度；从理论上推导了由仿射性质恢复度量性质的条件和将射影变形近似为仿射变形的条件,总结和归纳了常用的仿射不变量。针对城市地区低空倾斜影像几何变形大、视差断裂、高程不平滑的特点,提出了一种基于三阶矩阵的仿射不变特征提取算法,用于克服仿射变形提取影像中的不变特征,通过3组不同的影像类型进行实验,验证了算法的效果。针对低空倾斜影像重叠度大、冗余信息多的特点,从理论上分析了在匹配中筛选影像的必要性和可行性。针对多视影像,提出了一种多通道二进制评价模型用于在匹配的过程中实时的筛选影像,对局部影像的质量进行分析和排序,寻找最佳匹配像对,优化多片匹配模型的性能,筛选出“少而精”的影像对完成匹配的流程。针对城市地区建筑物影像存在视差断裂的特点,提出了一种由特征信息引导的自适应物方面元传播密集匹配算法,算法根据窗口范围内特征的视差边界自适应的调整面元大小,使面元在传播的过程中不越过视差边界,并且在平滑的表面面元面积扩大,在地势起伏变化的表面,面元变小,提高了算法的匹配效率。生成精确、可靠的三维点云模型。总结了地形场景海量点云数据下Delaunay构网算法的思路,基于目前大数据量构网普遍遇到的内存不足、CPU核心利用率低的情况,提出了一种基于并行多核心平台的算法,算法将OpenMP多线程编程技术,与流数据的并行构网相结合,并实时的释放运行内存,取得了不错的效果。实验基于标准航空影像和1+4倾斜影像两组数据,从视觉检测和定量评价两个方面来验证本文方案的实际效果。
[Abstract]:The high altitude aerial photography can only obtain the low resolution texture and the top texture information of the building because of the high flying altitude and the fixed space shooting attitude (vertical or small angle photography). It is difficult to meet the requirement of the three-dimensional model of the city building to the texture. The low altitude unmanned aerial vehicle combined with the newly developed multi-sensor integration technology. With the advantages of flexible, rapid and high resolution, it can be used to collect three-dimensional model information of urban buildings, which can meet the needs of building three-dimensional model of urban buildings. However, from the results of current market and scientific research units, the production efficiency of 3D landscape model is low and rough, and it is difficult to meet the needs of actual users. On the image of the sensor, one object point can be simultaneously imaging on multiple image images. Therefore, this is a contradiction of "more and less" data. On the one hand, the multi angle image overlaps high and the redundant information is large; on the other hand, the actual utilization rate of the image is low and the production of the 3D model is rough, which can not meet the needs of the users. A technical route of extracting low altitude tilted image feature information and guiding dense matching to generate 3D landscape model is worked out.
The contents of the thesis are as follows:
The basic principle of tilt photography is explained, and the difficulties in the stereo matching are analyzed. Then, the existing matching model is systematically analyzed, summarized and classified, and the existing feature corner detection operators based on the image square, the descriptor and the stereo matching algorithm based on the object are summarized. Constraint methods are summed up and summarized.
The geometric characteristics of projective transformation, affine transformation and similar transformation under the projective transformation group are systematically summarized, and the degrees of freedom between the invariants of each transformation and the transformation are summarized, and the conditions for restoring the metric properties from the affine properties and the conditions for the approximation of the projective deformation to be similar to the imitating deformation are derived from the theory, and the commonly used ones are summarized and summarized. An affine invariant. In view of the characteristics of the large geometric distortion of the low altitude tilt image in the urban area, the parallax fault and the elevation is not smooth, an affine invariant feature extraction algorithm based on the three order matrix is proposed to overcome the invariant features extracted from the affine deformation, and the experiment is carried out through 3 different types of image types to verify the effectiveness of the algorithm.
In view of the characteristics of low altitude tilt image overlapping and redundant information, the necessity and feasibility of screening images in matching are analyzed theoretically. A multi-channel binary evaluation model is proposed for real-time screening images in the matching process for multi view images, and the quality of local images is analyzed and sorted. The best match image pairs optimize the performance of the multi slice matching model, and select the "less" and "fine" images to complete the matching process.
In view of the characteristics of disparity in the image of buildings in urban areas, an adaptive element propagation dense matching algorithm is proposed, which is guided by feature information. The algorithm adaptively adjusts the size of the surface element according to the parallax boundary of the features within the window range, so that the surface element does not cross the parallax boundary in the propagation path and is smooth in the table. The area of the surface element is enlarged, and the surface element becomes smaller on the fluctuating surface, which improves the matching efficiency of the algorithm, and generates an accurate and reliable three-dimensional point cloud model.
This paper summarizes the idea of Delaunay network algorithm based on massive point cloud data of terrain scene. Based on the lack of memory and low utilization of CPU core, an algorithm based on parallel multi core platform is proposed. The algorithm combines OpenMP multithread programming technique with parallel network of stream data, and real-time The release of running memory has achieved good results.
The experiment is based on two sets of data of standard aerial image and 1+4 tilt image, and verifies the actual effect of the scheme from two aspects: visual inspection and quantitative evaluation.

【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：P231

【相似文献】