Three-dimensional Reconstruction Of Rotation Views

In the last few years, three-dimensional models have become more and more application fields, which range from engineering and architecture, to preservation of cultural heritage and entertainment. Following the growing demand of 3D models, the development of low cost acquisition systems has also become of key practical and theoretical value. Computer vision has provided important solutions to the low cost acquisition of 3D models, whose important task is to understand the structure of the object from views and one of outstanding achievements is to compute camera internal parameters, camera motion and recover 3D structure of object from multiple uncalibrated views.3D reconstruction from views has been one of the most popular research subjects in the field of computer vision, common approaches to 3D reconstruction are multiple views geometry, shape from shading, level sets methods and defocus images methods. The paper researches 3D reconstruction of both single uncalibrated view of surface of revolution and multiple uncalibrated views of single axes rotation motion. In the process of rotation, the trajectory of any one of points in the object is a circle, so the whole object will form a'virtual surface of revolution', therefore the paper research the two types. The main research works and achievements are outlined as follows:(1) Camera calibration is to compute camera internal parameters from views of object, and is one of the essential steps to 3D reconstruction from uncalibrated views, because projective reconstruction can be achieved and metric reconstruction can not be achieved if camera is uncalibrated. The paper focuses on camera calibration method based on coaxial circle, which method make use of the property of the image of absolute conic related to camera internal parameters, and cholesky decomposition of the image of absolute conic can obtain camera internal parameters. On the study of Colombo and others, a new formula to compute the image of absolute conic is derived and camera calibration is finished.(2) It's impossible to recover 3D structure of object from single uncalibrated view without any priori knowledge. The paper research 3D metric reconstruction from single uncalibrated view of surface of revolution, as the each cross section of surface of revolution is a circle, whose projection is a conic, and the nature provide enough priori knowledge for 3D metric reconstruction from single uncalibrated view of surface of revolution. The formula of conic computation is derived from planar homology and then the formula of conic back-projection is derived, finally 3D metric reconstruction of surface of revolution is achieved.(3) Three-dimensional reconstruction based on uncalibrated single-axis rotation image sequences is divided into two steps:camera calibration and 3D points acquisition. The similarity between the single-axis rotation and surface of revolution is found, and then camera calibration algorithm based on coaxial circle is introduced to camera calibration method of a single-axis rotation motion. Algorithm of structure from motion is introduced to compute 3D points coordinates, such that SIFT algorithm is used to compute feature points in each view and then nearest-neighbor algorithm is used to compute the matching points between adjacent images, finally linear triangle method is used to compute three-dimensional points of the object.