3D Geometry Reconstruction from Multiple Segmented Surface Descriptions Using Neuro-Fuzzy Similarity Measures

This paper presents a novel solution to the reconstruction of 3D geometry models from partial, segmented (2.5D or 3D) range views. First, the geometric fusion works entirely on sparse symbolic information, i.e. attributed surface graphs, rather than point data or triangulated meshes. Thus, new sensor data can always be integrated with an existing partial model available for symbolic action planning. Second, assumptions on automatic registration are weaker than those found in related work: the views need not be approximately calibrated, and no pre-existing knowledge of their overlap is needed. In order to find corresponding (redundant) surface features reliably even under high-noise and occlusion conditions we develop Neuro-Fuzzy similarity measures on surface descriptions. Third, we propose a reasonably complete prototype system including algorithms for merging sparse, reduced surface attributes, in particular boundaries. The experimental results from segmented range images of an indoor camera motion sequence demonstrate the ability to cope with unknown camera positions, low image resolution, large measurement and segmentation errors.