GraphReg: Dynamical Point Cloud Registration With Geometry-Aware Graph Signal Processing.

Area of Science:

• Computer Vision
• Geometric Computing
• 3D Data Processing

Background:

• 3D point cloud registration is crucial for 3D vision tasks.
• Existing physics-based methods often neglect surface geometry, leading to inaccuracies.
• Outlier sensitivity and computational inefficiency hinder current registration approaches.

Purpose of the Study:

• To develop a high-accuracy, efficient, and robust 3D point cloud registration method.
• To integrate surface geometry into physics-based registration using rigid-body dynamics.
• To improve robustness against outliers and accelerate the registration process.

Main Methods:

• Leveraging Graph Signal Processing (GSP) for point response intensity.
• Employing Median Absolute Deviation (MAD) and the X84 principle for robust outlier handling.
• Utilizing a novel geometric invariant for force modeling and Adaptive Simulated Annealing (ASA) for optimization.

Main Results:

• The proposed method achieves superior accuracy and robustness compared to state-of-the-art approaches.
• Demonstrated suitability for large-scale point cloud registration from range scanners and LiDAR.
• Exhibits significantly faster processing times and enhanced stability.

Conclusions:

• The geometry-aware rigid-body dynamics approach offers a significant advancement in 3D point cloud registration.
• The method provides a robust, efficient, and accurate solution for diverse 3D vision applications.
• Publicly available implementation facilitates further research and development.