A Robust Multiresolution Registration Approach

The task of registering several range images taken from different viewpoints into a single coordinate system is usually divided into two steps: first, finding a rough estimate of the searched transformation on the base of reduced parts of the data sets, and second, finding the precise transformation with another method on the base of the complete information given by the data sets. Using different approaches for these registration steps has no fundamental but only practical reasons: for the fine tuning step it exists an easily implementable algorithm whose efficiency and precision has been proven in many experiments, but which is not applicable for finding an initial rough estimate of the searched transformation. We present a multiresolution approach to the registration problem that has the potential to combine these two registration steps and is based on hierarchical Hough methods and local frames defined in each surface point.