Based On 3D Reconstruction Research Of Depth Image

The mass point cloud obtained by the ground laser radar scanning technology can express entity object accurately. But because of the point cloud data have scattered disorderly features three-dimensional point cloud model reconstruction has been a difficult problem. In particular, with the growing of the number of point cloud data currently acquired and the improving of the accuracy of point cloud constantly, the demand for refined model has become more and more urgent. Therefore, it is an important issue currently how to reconstruct three-dimensional model fast, precisely and efficiently.At present, the reconstruction of the point cloud model mostly expressed by triangulated irregular network. It has advantages of wide universality, and it can be used to express fine geometric textures. But there still exist some defects, such as less semantic, unable to extract geometric characteristic parameters, need to store large amount of data and the inconvenience of the using of the model subsequently. However, fitting voxel model by point cloud data could make up for the deficiencies above well. This method is often used for smooth geometric objects such as pipes, walls, etc, and the defect of this method is that it can’t express the object with prominent surface geometry texture. The hybrid model based on the above two can complement each other, and it is the development trend in the design of spatial data model. However, some model need consume so big storage resources that the efficiency is not high in practical application.We expanded the concepts of depth image in three-dimensional, and then combined the advantages of triangulated irregular network modeling and voxel modeling. Finally, we proposed the depth image modeling based on geometric primitives. In this paper, the research is conducted around the geometric primitives fitting and the depth image generation, taking terrestrial laser point cloud data as the research object.We use plane, sphere and cylinder geometric model as the research objects in geometric primitives fitting phase. This paper adopts two strategies according to different applications.One is using region growing method to fit geometric primitives based on seed point, the other is fitting based on segmentation point cloud integrally. For the sphere fitting, we provide a robust estimation of iteration method to improve the fitting accuracy. For the cylindrical model whose model parameters are linearly related, this paper adopts the nonlinear optimization algorithm named Levenberg-Marquardt for iterative computations. By comparing the initial with different calculation methods, we present a robust method to extract the initial for cylinder.In depth image generation, the paper firstly transformed the local coordinate system depending on the individual characteristics of geometric primitives. Secondly, we divided grids on the surface of a geometric model and interpolated at grid node. Finally, we reconstructed the three-dimensional model. In the process of reconstructing plane depth image model, we use OBB bounding box as the basic bounding box for the grid division, so that thegrid can be divided in accordance with the direction of the diameter of point cloud. For the reconstruction of the holes exist in the model, this paper presents an approach which can automatically fill the hole based on the detection of holes. In the process of spherical depth image model, this paper uses the quaternary triangulation(QTM) technology for grid division after comparing the existing spherical grid division methods. The method can divide the spherical model uniformly, so it is possible to obtain a uniform texture spherical depth image.In the process of reconstructing cylinder depth image model, we rapidly calculate the upper and lower vertices of the cylinder through the cylinder model parameter extracted using the method of point to linear projection. Then we launched the surface of the cylindrical model on a plane surface for grid division. To generate a seamless cylindrical depth image, this paper interpolated the expanded edge by increasing the width of the plane. Finally, we generate the depth image taking the planar, spherical, cylindrical as the reference datum.To verify the feasibility of the algorithm, using C/C++ language developed the algorithm with the Visual Studio 2013 development platform based on the CLR framework on the above the achievements research of algorithm. Finally, the article verified the practicability and effectiveness of the algorithm in this paper, using the point cloud data of the Yunlong stone carving and the ding dedicated to his mother amyl as the data of experiment which scanning by scanner in handheld and articulated arm scanner.