| 3D model of ground and objects plays a key role and is widely used in the field of geosciences and others, so, acquisition of 3D model of ground and objects, is a hot topic in the field of photogrammetry and computer vision. In the field of photogrammetry, the 3D model of the ground is represented by Digital Surface Model (DSM) and Digital Elevation Model (DEM), which are the basic data of Geographic Information System (GIS). It has important uses in land management and planning, route design, power, transportation, geology, hydrology and urban planning and other fields. In the field of computer vision,3D models of objects and scenes have an indispensable role in the fields of digital archives, game and entertainment industries.Image-based 3D reconstruction is a key technique to recover the shape of ground and objects, which is the reverse process of imaging. The imaging process consists of two parts: geometric imaging and radiometric imaging. Multi-view stereo methods and shading-based shape recovery methods reconstruct the 3D shape of ground and objects according to the geometric part and radiometric part, respectively. There are a lot of researches on multi-view stereo and shading-based shape recovery, and remarkable progress has been made. The former mainly developed the local matching method and the global matching method, while the latter mainly developed shape from shading and photometric stereo vision. There are also many studies on combination of the two, but they are only limited to recover the 3D shape of the surface of a single material such as lunar surface, gypsum and sculpture. There are few 3D reconstruction methods that extend it to the earth's surface and natural objects.In this paper, a dense matching technique framework based on geometric and radiometric integral imaging model suitable for geoscience application is developed, which combines the geometric and radiometric properties of the image. Under this framework, the 3D shape recovery of the earth surface and natural objects is realized. In particular, the main research work of this paper is as follows:(1) Dense matching strategy based on geometric and radiometric integral imaging model is studied. The geometric imaging model and the radiometric imaging model are unified, and the geometric and radiometric integral imaging model is established. The correspondence between the points on the object surface and the position and brightness of the pixels on the image is revealed. And the essence of multi-view stereo technique and shading-based shape recovery technique is analyzed. On this basis, dense matching strategy based on geometric and radiometric integral imaging model is proposed. Considering the constraints of stereo geometric and radiometric property, the energy functional is constructed which is consistent with photo-consistency, geometric smoothness and radiometirc smoothness, and a complete solution strategy of the energy functional is proposed.(2) 3D reconstruction method based on multi-metric semi-global matching is studied. Firstly, the global optimization strategy of selecting the best pairing image from the neighborhood candidate image set based on the Markov random field theory is studied to automatically select the image pairs. Secondly, according to the short of the classical semi-global matching on efficiency and robustness, the multi-metric semi-global matching (mSGM) algorithm is proposed which is improved in the aspects of the choice of penalty coefficient, the choice of similarity measure, the adaptive adjustment of disparity range, the calculation of matching confidence and the improvement of image radiation quality. This algorithm combines the advantages of Census and mutual information similarity measure, adopts the pyramid image matching strategy, takes the upper disparity result as the initial value and adjusts the disparity search range of the lower layer adaptively to the terrain. It is robust, reliable and efficient and edge-preserving. The experiment results show that mSGM is better than SURE in terms of completeness and fineness. Finally, the method of disparity fusion and surface reconstruction is studied.(3) A global matching method for multi-view images based on bilateral filtering is studied. In this paper, the global energy function is constructed on the imaging light cone space. The multi-view photo-consistency is taken as the matching cost, and the geometric smoothness and radiometric smoothness of adjacent points are taken into account. A two-step updating model based on bilateral filtering is used to solve the energy functional The solution of the traditional partial differential equation (PDE) is transformed into two steps. One step is to solve the PDE based on the data item, and the other is equivalent to bilateral filtering with multiple-cues.3D surface results of mSGM is used as input, and the final 3D surface is obtained through global optimization.In this paper, linear array satellite images, conventional aerial images, oblique aerial images and close-range images are used to carry out dense matching experiments, to validate the applicability and effectiveness of the algorithm. In the aspect of linear array satellite imagery, DSM was generated by dense matching of ZY-3 satellite images of Yuntai Mountain in Zhejiang Province, and the DEM product of 1:10000 scale was used as reference for quality analysis and precision evaluation. In terms of quality, DSM details were close to the reference DEM; in accuracy, DSM elevation accuracy satisfied the precision requirements of national's 1:50,000 scale DEM products. In conventional aerial imagery, Vaihingen and Munchen datasets are used to carry out dense matching experiments. The matching results in this paper are superior to Pix4Dmapper and PhotoScan in detail and anti-repetition texture. For oblique aerial imagery, matching experiments were carried out with urban oblique images in Tongchuan City, Shanxi Province, and the result was better than that of PhotoScan. In the close-range imagery, the matching experiment was carried out by using smart phone images of the stone lions statue of Friendship Square in Wuhan University. The matching effect was comparable with that of PhotoScan, which was better than that of PMVS. |
PDF Full Text Request