Three dimensional reconstruction as an automaic modeling system
Xiaobu Yuan · ERA: Education and Research Archive (University of Alberta) · 1993
Research in three-dimensional (3D) scene reconstruction is aimed at developing techniques that enable a vision system to acquire 3D representations of its environment from sensory input. By introducing a new data structure, namely oriented surface models, this thesis presents a modeling system that is able not only to accumulate observed features from different views automatically, but also to recognize objects and retrieve processed information. The oriented surface data structure is a modified surface modeling method. It describes scene objects with locally defined surface primitives, and specifies surface geometric locations and topographical relations with homogeneous transformations. Its viewpoint independent object representation makes reconstruction systems capable of accumulating and identifying spatial features during model building and model retrieving. And its weighted vector chains provide a useful clue for a system to automatically check the closure of boundary surfaces and to estimate optimal viewing directions. The thesis briefly surveys issues of 3D reconstruction in the first several chapters, which includes techniques to extract and describe spatial features, build object models, and recognize objects. Following it are the description of the new data structure and a discussion of how it is applied to object reconstruction and recognition. The system has been simulated on SUN workstations with a group of graphics objects. The system's implementation and analysis, as well as experiments with real range images are discussed.