A Robust Point Cloud Registration Method Using Geometrical Information and Probabilistic Models for Robot-Assisted Bronchoscopy Navigation

Point cloud registration in robot-assisted surgery is a fundamental procedure, which is directly related to navigation accuracy and patient safety. However, preoperative reconstructions furnish dense point cloud representations, while intraoperative point clouds only yield partial point sets. In addition, the intraoperative measurements from diverse sensor modalities are often disturbed by substantial noise, thereby posing a challenging partial-to-full registration problem. Therefore, we propose a robust registration framework underpinned by geometric information and probabilistic modeling to overcome these limitations and align intraoperative measurements with preoperative volumetric models. First, an initial rigid transformation is obtained by matching global geometric features. Second, we jointly characterize spatial deviation in both point cloud via Gaussian mixtures and models orientational consistency via the cosine of the corresponding normal vector angle to simultaneously drive geometric and directional alignment. Next, we recast registration as a maximum likelihood estimation problem and embed a Kullback–Leibler (KL) divergence regularizer into the posterior update step to enhance robustness and convergence stability. A robot-assisted bronchoscopy platform is established, on which simulation and phantom experiments are conducted to verify the comprehensive performance. The experimental results quantitatively and qualitatively demonstrate that our framework can obtain accurate and robust registration performance.