Multilevel Optimization for Registration of Deformable Point Clouds.

Area of Science:

• Computer Vision
• Geometric Processing
• 3D Data Analysis

Background:

• Point cloud registration is crucial for 3D data analysis.
• Object deformation presents a significant challenge in accurate point cloud registration.
• Existing methods struggle with dynamic changes in object shape and location.

Purpose of the Study:

• To develop a robust point cloud registration method capable of handling significant object deformation.
• To leverage both local and global geometric features for improved registration accuracy.
• To introduce a novel energy function integrating multi-level geometric information.

Main Methods:

• A new energy function combining local, semi-local, and global geometric terms was defined.
• Transformation parameters were estimated using a closed-form solution for local geometry.
• Nonlinear least squares minimization via Gauss-Newton optimization was employed for global geometry.
• Block coordinate descent was used to optimize the total energy function.

Main Results:

• The proposed method demonstrated superior performance in registering deformed point clouds.
• Experimental results on three public datasets show significant improvements over state-of-the-art algorithms.
• The integration of multi-level geometric attributes proved effective in handling deformation.

Conclusions:

• The developed approach offers a robust solution for point cloud registration under deformation.
• Combining local and global geometric cues enhances registration accuracy for dynamic objects.
• This method advances the field of 3D data registration, particularly for animated or deforming objects.