Shape Registration Using Deformable Self-Organizing Feature Maps

摘要

A novel approach to matching freeform surfaces for shape registration and object recognition is described in this paper. The proposed method builds a surface mesh of the underlying object geometry by iteratively deforming the nodal lattice of a spherical self-organizing feature map (SOFM) to "best" fit the measured 3D coordinate data. The final topology of the deformed mesh is, therefore, equivalent to the original lattice of the SOFM. Each node in the final mesh represents a cluster of coordinate points that lie in close spatial proximity in the input data space. In this way, closed surfaces with identical node topologies are created from different data sets. Information about node connectivity is then extracted from the ordered lattice and used to determine local surface features for correspondence matching. Based on the matched nodes, rigid body transformations between the original data sets can be determined. The shape registration algorithm enables comparisons to be made between different sized data sets or data acquired from similar freeform objects with arbitrary pose. The method is illustrated using measured coordinate data from three objects with complex freeform surface geometry.