Generating three-dimensional surface models of solid objects from multiple projections
Michael Potmesil · 1982
A recurring problem in computer vision and related fields is that of generating computer models of physical objects. This thesis presents a method for constructing such models in the form of three-dimensional surface or volume descriptions. The surface models are composed of curved, topologically rectangular, parametric patches. The data required to define these patches are obtained from multiple photographic images of the object illuminated by a rectangular pattern of lines. The projection of the pattern on the surface of the object traces curves which define the boundaries of the patches. The 3D description of the patches is reconstructed by photogrammetric techniques from two or more images of the projected pattern. A calibration stand, in which the object is placed, permits determination of the camera geometry directly from image data. This method generates 3D surface descriptions of only those parts of the object that are illuminated by the projected pattern, and also are visible in at least two images. A complete model of the object is obtained by repeating this reconstruction process for various arbitrary orientations of the object until descriptions covering the entire surface have been obtained. Since each description is defined in its own 3D object space and 2D parameter space, a surface-matching procedure is developed to register spatially the surface descriptions in a common object space. This procedure searches for a 3D rigid transformation of the surface descriptions which minimizes their shape difference. Once the surface descriptions are in the same object space, they are also merged into a common parameter space. This match-and-merge process is iteratively repeated for pairs of surface descriptions until a complete model of the object is assembled.