Robust CPD Algorithm for Non-Rigid Point Set Registration Based on Structure Information.

Area of Science:

• Computer Vision
• Geometric Processing
• Machine Learning

Background:

• Coherent Point Drift (CPD) is a popular algorithm for point set registration.
• CPD's limitations include ignoring point neighborhood structures and requiring manual outlier ratio assignment, reducing robustness to data degradation.

Purpose of the Study:

• To enhance the Coherent Point Drift algorithm for improved robustness.
• To address CPD's limitations in handling neighborhood structure and outlier ratios.

Main Methods:

• Incorporated a structure descriptor (e.g., shape context) for auxiliary correspondence calculation.
• Adjusted Gaussian Mixture Model (GMM) component proportions based on point similarity.
• Integrated outlier ratio calculation within the Expectation-Maximization (EM) framework for automatic optimization.

Main Results:

• The proposed method demonstrates superior robustness compared to standard CPD.
• Experimental results show improved performance on synthetic and real data with deformation, noise, occlusion, and outliers.
• Outperforms other state-of-the-art registration algorithms in challenging conditions.

Conclusions:

• The improved CPD algorithm effectively overcomes the limitations of the original method.
• The integration of structural information and automated outlier handling significantly enhances registration accuracy and robustness.
• This approach offers a more reliable solution for point set registration in the presence of significant data degradation.