Object shape modelling from multiple range images by matching signed distance fields

Modelling object shapes from multiple range images requires three procedures: removal of measurement errors, registering input shapes and integrating them as a shape representation. We propose a unified framework in which these procedures can be solved simultaneously. Discrete samples of the signed distance field (SDF) from the object surface are used as the shape representation. If the data shapes are registered correctly, the SDF should match in the common coordinate system. The data shapes are first integrated by averaging the data SDF assuming that they are roughly pre-registered. The data shapes are registered to the integrated shape by matching the SDF. Integration and registration are iterated until the input shapes are registered to the integrated shape. Weighting values are controlled to reject outliers caused by measurement errors and wrong correspondences. The proposed method does not suffer from cumulative pairwise registration errors because all data shapes are registered to an integrated shape. A polygonal surface model is generated from the integrated SDF. The method was tested on synthetic and real range images, and multiresolution results are also presented.