The Large-scale Scene 3D Reconstruction Algorithm Based On Global Positioning System

The large-scale scene 3D reconstruction has become one of the trends in the computer vision research in the future. With the significance in the field of computer vision, the large-scale scene 3D reconstruction also has a widespread application in other domains, such as surveying and mapping and urban planning. However, there is still lack of a proper method to solve the large-scale scene 3D reconstruction problem. Since the limitation of the ordinary camera’s view, the traditional reconstruction algorithm which is based on the general camera model suffers from the low efficiency and the discontinuity of the reconstruction process. And this problem results in the inapplicability of the current algorithm to the large-scale scene 3D reconstruction project. Meanwhile, the large-scale scene 3D reconstruction process contains such problems that the accumulative error is huge and the real world scale cannot be introduced and the calculation is time-consuming. In terms of these issues, it is necessary to acquire an improved algorithm to satisfy the needs for the large-scale scene 3D reconstruction.In this thesis, some methods of panoramic 3D reconstruction are deeply studied, and an improved large-scale scene 3D reconstruction algorithm based on Global Positional System is proposed. A self-built multi-sensor acquisition system is used to get panoramic images and position information, and a sparse point cloud structure of the large-scale scene is reconstructed finally.Firstly, according to the complex terrain and traffic conditions in the urban, a multi-sensor panoramic image acquisition system is designed, which is used for providing high-class image and position information for the following reconstruction work.Secondly, on the basic of the definition and property of spherical camera model, a panoramic 3D reconstruction algorithm based on spherical camera model is designed. This section contains the spherical transformation of the ordinary SIFT feature extraction and matching algorithm and the 3D recovery algorithm based on general camera model. This transformation makes the 3D reconstruction satisfy the panoramic images well.Thirdly, in terms of the scale and accumulative error problem in the large-scale scene 3D reconstruction, the Bundle Adjustment algorithm in reconstruction process is deeply studied. An improved Bundle Adjustment algorithm with weight and an improved Bundle Adjustment with position constraint are achieved respectively, which are compared in the end.Finally, with the relative knowledge of the parallel computing and the graphic processor, an acceleration method for the large-scale scene 3D reconstruction is discussed. The improved Bundle Adjustment algorithm is parallel programmed and accomplished in the GPU, which is aimed to reduce the waste of time.