Spatially-constrained Keypoint Matching for Efficient Statistical Shape Modelling

Abstract

Statistical shape models (SSMs) allow the compact description of the variability of object shapes within a given sample set. They are commonly used in medical imaging to model and analyse the shape of anatomical structures such as organs. The generation of a SSM mainly consists of the calculation of the average shape and the main directions of variation of the data set. Usually, structured point clouds are used as shape representations. A crucial step in the calculation of the average shape of the data set represents the transformation of the objects into a common reference space in order to average coordinates of corresponding points. When using unstructured point clouds without explicitly defined landmarks, the matching of correspondences remains a challenge. We propose a novel solution for a spatially-constrained keypoint matching (FPFH++). It is based on a first attempt of a feature-based registration using Fast Point Feature Histograms (FPFH) compared to a baseline approach utilizing Coherent Point Drift (CPD).

Keywords

• statistical shape models
• registration
• feature extraction