Description of complex objects using multiple range images

Abstract

Building computer models of existing 3-D objects from sensor data is an interesting yet very difficult problem. This research has applications in computer vision research as well as in industry for such purposes as inspection and reverse-CAD. The acquired models can also be used for robotic planning and manipulation. We identify three key steps in the model construction process, which are, (1) data acquisition, (2) data integration and (3) description. We use range images of an object from multiple viewpoints as input to our system since range images provide direct 3-D measurements of the object surface. To integrate the range image views, we develop a new algorithm for range image registration, which can find an accurate rigid 3-D transformation between a pair of range images given coarse transformation estimates. Our formalization of the registration problem is less constrained than the commonly used schemes based on point-correspondence, resulting in a much faster convergence. A global registration scheme is also presented to integrate all the range image views for the description task. Our description scheme can achieve both integration and description at the same time. We use a dynamic triangulated mesh with spring attachment between the vertices as the shape model. We introduce an inflation force inside the balloon-like model initially placed inside the object with a spherical shape. The inflation force expands and deforms the mesh until it reaches the object surface. A dynamic mesh subdivision and a local adjustment scheme are also incorporated so that the mesh eventually fits the entire surface of the object, accomplishing the mapping from a sphere to any complex object surface. The novel point of the approach is that only inflation force is introduced during the mesh deformation process. The problem with other deformable model and dynamic mesh methods, which is the reliance on a good initial guess for the global optimum, is thus no longer an issue. Experimental results for modeling both simple and complex objects from real range images are given. (Copies available exclusively from Micrographics Department, Doheny Library, USC, Los Angeles, CA90089-0182).