Recently the use of three-dimensional computer models has greatly increased, in part because of the availability of fast, inexpensive graphics hardware and technologies such as VRML-ready Internet browsers. These models are often of existing objects and are typically built by hand using CAD software, an error-prone and labor-intensive process. This thesis investigates methods by which these models may be automatically acquired, processed, and utilized by using range data. An incremental modeling method is described that builds accurate solid models of objects from multiple range images. A hybrid of surface mesh and volumetric representations is used to create a “water-tight” 3-D model at each step of the modeling process, allowing models to be built from a small number of range images. The method is able to model scenes consisting of multiple, disconnected parts without imposing restrictions on their topology. The resulting models retain information that denotes each surface element as properly acquired or requiring additional sensing. An important part of the model acquisition process is the determination of the next sensor viewpoint. We introduce a planning method that computes visibility for model surfaces to determine occlusion-free sensor positions that ensure that model fidelity is improved. These sensor positions are computed in continuous space, allowing a more complete solution then planning methods that rely on spatial discretization. These two processes are combined in a system capable of acquiring models of a wide variety of shapes and scenes. Examples are shown using objects with features such as holes and disconnected parts, as well as scenes with large self-occlusions. Applications of this work include graphics, manufacturing, robot navigation and architectural site modeling.
Recommended citation: Michael Reed, Solid Model Acquisition from Range Imagery, Ph.D Thesis, Dept. of Computer Science, Columbia University, 1998.)