Fast Descriptors and Correspondence Propagation for Robust Global Point Cloud Registration

Area of Science:

• Computer Vision
• Geometric Computing
• 3D Data Processing

Background:

• Point cloud registration is crucial for 3D reconstruction and analysis.
• Existing methods struggle with low-precision matches, noise, and varying point cloud overlaps.
• Robust and efficient global registration remains a significant challenge.

Purpose of the Study:

• To develop a robust and efficient global approach for point cloud registration.
• To address the challenges of low-precision initial matches and varying alignment conditions.
• To provide a method applicable to diverse 3D scanning scenarios.

Main Methods:

• Utilizing multi-scale eigenvalues and normals for fast local structure descriptors.
• Employing a correspondence propagation mechanism to aggregate initial seed matches.
• Computing multiple transformations and selecting the best coarse alignment using a quality function.
• Refining the alignment with the trimmed iterative closest point (TICP) algorithm.

Main Results:

• The proposed method achieves robust global registration for point clouds with significant or limited overlaps.
• It demonstrates high efficiency and resilience to noise.
• Performance is superior to traditional descriptor-based and other global registration algorithms.
• Successful application in large-scale reconstruction and registration of low-resolution Kinect scans.

Conclusions:

• The developed approach offers a reliable and efficient solution for point cloud registration.
• Its robustness to noise and ability to handle diverse overlaps make it suitable for real-world applications.
• The method advances the field of 3D data processing and reconstruction.