Abstract
In this paper, we formulate a probabilistic point set matching problem under variational Bayesian framework and propose an iterative algorithm in which the posteriors of parameters are updated in sequence until a local optimum is reached. This variational Bayesian registration approach explicitly accounts for the matching uncertainty in terms of the parameters and is thus less prone to local optima. Furthermore, the anisotropic covariance is assumed on each individual component of Gaussian mixtures and is estimated by the iterative approximate process. Experimental results show that the combination of variational Bayesian approach with Gaussian mixtures obtains favorable performance with respect to the accuracy and the robustness in comparison with other registration algorithms.