Generation of geometric model by registration and integration of multiple range images

We propose a method to construct a 3-D geometric shape model by registering and integrating multiple range images under a unified framework. We use 3-D points around the object surface as the correspondence key, because the closest data point from an arbitrary 3-D point should be matched if the data are registered correctly with some exceptional cases. We assume that the input shapes are roughly registered in advance, and then we solve the problem by iterative loops that consist of establishing correspondence by searching closest data points, integration of corresponded data points, and registration of each data to the integrated shape. The error function is defined by the weighted variances of the signed distance field from the surface. The weighting values are determined to trim out wrong correspondences typically caused by different sampling and measurement error. We do not need to consider the cumulative pairwise registration error because each data is registered to the integrated shape. The integrated shape information is directly used to generate a surface model. We tested the method on synthetic and real range images.