Robust non-rigid point registration based on feature-dependant finite mixture model

• A feature-dependant finite mixture model (FDMM) is proposed to fit the non-rigid point set registration.
• An improved shape descriptor is used to define the local shape feature of a point.
• Two point sets are updated by a bidirectional mechanism.
• The new algorithm improves the robustness and accuracy with heavy outliers comparing to traditional methods.

In previous works on point registration based on finite mixture model, the correspondence probability is often determined by exploiting global relationship in the point set instead of considering the local point distribution. That results in a simplified registration model. In this paper a feature-dependant finite mixture model (FDMM) is proposed. In particular, an improved descriptor is introduced to describe the local feature of a point. Consequently, a priori density function is formulated for the mixture weights. The unknown parameters of FDMM are computed by maximizing a posteriori (MAP) estimation. Moreover, a bidirectional expectation–maximization (EM) process is introduced to update both point sets in contrast to traditional methods. The performance of our method is demonstrated and validated with carefully designed synthetic data and real data, showing that the proposed method can improve the robustness and accuracy as compared to the traditional registration techniques.