Hierarchical Optimization of 3D Point Cloud Registration

Area of Science:

• Robotics
• Computer Vision
• Computational Geometry

Background:

• Rigid registration of 3D point clouds is crucial for robotics and computer vision.
• Iterative Closest Point (ICP) algorithms are widely used but suffer from sensitivity to outliers, initial pose, and computational inefficiency.
• Existing ICP variants often compromise accuracy or increase complexity to address these limitations.

Purpose of the Study:

• To propose a novel hierarchical optimization approach for 3D point cloud registration.
• To develop an algorithm that overcomes the limitations of traditional ICP methods, specifically regarding outlier sensitivity and computational efficiency.
• To enhance the accuracy and robustness of 3D point cloud registration.

Main Methods:

• A hierarchical optimization approach combining an improved voxel filter with Multi-Scale Voxelized Generalized-ICP (MVGICP).
• Outlier filtering and downsampling are achieved by integrating traditional voxel sampling with point density.
• MVGICP utilizes multi-scale iteration and avoids closest point computation to prevent local minima and optimize efficiency.

Main Results:

• The proposed improved voxel filter effectively handles outliers and downsamples point clouds.
• MVGICP demonstrates superior performance in outlier filtering compared to existing algorithms.
• Experimental results show enhanced registration performance and computational efficiency of the proposed MVGICP algorithm.

Conclusions:

• The proposed hierarchical optimization approach, particularly MVGICP, offers a significant improvement over current 3D point cloud registration methods.
• The algorithm effectively addresses the challenges of outlier rejection and computational cost.
• MVGICP provides a more robust and efficient solution for 3D point cloud registration in robotics and computer vision applications.