Modelling object shapes from multiple range images requires three tasks: removal of measurement errors, registration of input images and integration of them as a shape representation. We propose a method that enables combining these tasks in a unified framework assuming that the input range images are roughly pre-registered. In this method, control points are taken on 3-D lattice points in the common integration coordinate system. If the data shapes are registered to the common coordinate system correctly, the closest data point from each control point should match. In addition to the closest point coordinates, the surface normal forms a local linear approximation of the signed distance field (SDF) in the neighbourhood of each control point. The data shapes are integrated by averaging the SDFs at each control point, and the total error is defined by the sum of weighted variances of the SDFs. Each data shape is registered to the integrated shape reducing the SDF error, and the procedures of integration and registration are applied alternately until they are converged. The weighting values are determined according to the compatibility of SDFs to reject outliers caused by measurement errors or false correspondences. The method does not suffer from cumulative pairwise registration errors because all data shapes are registered to an integrated shape. The integrated shape is directly used to generate a surface model. The method was tested on synthetic and real range images, and multiresolution results are also presented.
