This dissertation is on the 3D representation and reconstruction, normalization and recognition of 3D objects based on multiple views. To begin with, the general research situation in the field of computer vision is reviewed, and the fundamental theoretical frameworks of this field are summarized. The research background and the main topics of this dissertation are also discussed. At the same time, the methods for 3D representation are reviewed in detail with a brief introduction to the methods of 3D modeling. As to the 3D objects representation and reconstruction, the generalized linear octree representation is presented, and the formula for computing the generalized views are derived. Furthermore, the reconstruction algorithms for linear octree are extended to the generalized coordinate system to fulfill the generalized 3D reconstruction. With the introduction of the generalized representation, the orthogonal condition of the views is no longer a limitation. Thus on the one hand the 3D reconstruction can be less constrained and more flexible; on the other hand the views are no longer restricted to a limited number, and by increasing the views 3D objects can be reconstructed more precisely. For the computation of generalized views, two discrete methods for linear transformations of digital images are presented: point-governed region method and the method based on neighborhood features. The former consists of point labeling algorithm, interior point labeling algorithm, area element labeling algorithm and area element computing algorithm. The latter is an extension of Cheng's algorithm, and it can be used under general linear transformation instead of being suitable only to special cases of linear transformations of translation, scaling and rotation. Experiment results show that these two methods, used under general linear transformation, have the advantages of keeping the connectivity of the transformed images, eliminating superfluous holes of them and thus improving the transformation quality of these images. In the normalization of spatial orientation of 3D objects, a new method is presented. First, the problem of determining the uniqueness of principal axes of an object is analyzed theoretically. Then the criteria for determining the unique principal axes are derived. Based on 3D moments, the criteria are more stable and reliable compared with the method employing only single point to determine the unique principal axes. Moreover, no matter the centroid is inside the bounding surfaces of the object or not, the normalization result remains unchanged. As regards the normalization and recognition of 3D objects, an integrating approach which combines principal axes method (PAM) and moment method (MM) is presented. PAM and MM are the two major ones for 3D object normalization and recognition. It is easy to use the former to normalize the orientations of 3D objects but this method can not be u
修改评论