Non-Rigid 3D Point Set Registration With Reliable Hybrid Mixture Model

Abstract

Non-rigid point set registration is an essential technique in fields such as robotics, computer vision, image-guided surgery, and augmented reality. However, effectively addressing non-rigid deformations and handling noise interference remains a significant challenge in this context. This paper introduces a novel hybrid mixture model (HMM), combining a Gaussian mixture model (GMM) to describe positional features and a Fisher mixture model (FMM) to represent orientation features. This hybrid mixture model offers a solution to the non-rigid point set registration (PSR) problem. The proposed approach iteratively optimizes model parameters through the maximum expectation (EM) algorithm. It comprehensively explores registration performance, accounting for non-rigid deformations amidst isotropic and anisotropic positional noise. The proposed approach further incorporates reliable normal vectors for evaluating orientation features. Extensive experiments on non-rigid PSR demonstrates the improvements our algorithm offers in terms of both robustness and accuracy.