<p>One persistent challenge in mixed reality (MR) real-world applications is to reliably achieve accurate, unsupervised cross-source point cloud registration between real-world objects and their corresponding digital twin counterparts. Widely adopted traditional algorithms and recent deep learning-based approaches typically deliver unsatisfactory robustness and accuracy, limited by incorrect point pair correspondence caused on different distortions from MR device sensor sources, sampling resolution discrepancies, scarce real-world training datasets with ground truth annotations, human errors and high variance in initial pose estimation in real-world freehand MR scan scenarios. We propose a statistical registration method that can be recursively applied to point cloud created by different 3D senor sources in different resolution levels, using Otsu statistical clustering and statistical Wasserstein distance to remove point pair correspondence outlier and improve the overall confidence level in point pair correspondence, which is the loss function can iteratively improve registration affine matrix accuracy. Additionally, we propose using point pair distance’s global statistical distribution and its statistical distance to optima registration as a better registration quality metric over commonly used local single value metrics like fitness and RMS. The latter often fail to represent registration accuracy adequately as it is strongly dependent on confidence level of local pair-wise correspondence. Results demonstrate the improvement and effectiveness of our method through comparative analysis with existing approaches across benchmark datasets featuring varying levels of noise, distortion and 6-DoF resolution challenges.</p>