Non-Rigid Point Set Registration by Preserving Global and Local Structures

Area of Science:

• Computer Vision
• Computational Geometry
• Machine Learning

Background:

• Traditional point registration often uses Gaussian mixture models, focusing on global relationships.
• Non-rigid shapes require consideration of local structures for accurate point correspondence.
• Existing methods may struggle with data exhibiting significant distortions.

Purpose of the Study:

• To develop a robust point registration method that integrates both global and local shape information.
• To improve point correspondence accuracy for non-rigid shapes under various distortions.
• To enhance the performance of registration algorithms on degraded datasets.

Main Methods:

• Formulating point registration as a mixture density estimation problem.
• Utilizing local features like shape context for mixture model membership probabilities.
• Employing a reproducing kernel Hilbert space for transformation specification.
• Implementing a sparse approximation for computational efficiency.

Main Results:

• The proposed method effectively preserves both global and local structures during point set matching.
• Demonstrated robustness against deformation, noise, outliers, rotation, and occlusion.
• Significantly outperformed state-of-the-art methods, particularly on corrupted data.
• Achieved fast implementation through sparse approximation.

Conclusions:

• The novel mixture density estimation approach offers superior point registration performance.
• Integrating local features enhances robustness and accuracy in non-rigid shape matching.
• This method provides a significant advancement for handling challenging, degraded point cloud data.