Non-rigid point set registration via global and local constraints

Non-rigid point set registration is often encountered in meical image processing, pattern recognition, and computer vision. This paper presents a new method for non-rigid point set registration that can be used to recover the underlying coherent spatial mapping (CSM). Firstly, putative correspondences between two point sets are established by using feature descriptors. Secondly, each point is expressed as a weighted sum of several nearest neighbors and the same relation holds after the transformation. Then, this local geometrical constraint is combined with the global model, and the transformation problem is solved by minimizing an error function. These two steps of recovering point correspondences and transformation are performed iteratively to obtained a promising result. Extensive experiments on various synthetic and real data demonstrate that the proposed approach is robust and outperforms the state-of-the-art methods.