Multi-View Laser Point Cloud Global Registration for a Single Object.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Global registration of multi-view laser point clouds is crucial for 3D reconstruction of moving objects.
• Challenges include severe noise, density variations, and overlap ratios in point cloud data.

Purpose of the Study:

• To propose a robust global registration method for multi-view laser point clouds.
• To enhance the accuracy and stability of 3D reconstruction for moving objects.

Main Methods:

• Utilized spatial rasterization for feature extraction and loop-closure detection.
• Employed low-rank sparse decomposition with a robust objective function solved via the Augmented Lagrange method.
• Designed iterative termination conditions based on single-object global registration prior conditions.

Main Results:

• Achieved loop-closure detection accuracy exceeding 90% with robust parameter performance.
• Demonstrated superior global registration accuracy (<0.05 rad) compared to existing methods, especially with pairwise registration errors below 0.1 rad.
• Validated method's effectiveness and stability through simulations and real-world experiments.

Conclusions:

• The proposed low-rank sparse decomposition method offers a robust and accurate solution for global registration of challenging multi-view laser point clouds.
• The technique significantly improves 3D reconstruction of moving objects, outperforming current methods in accuracy and stability.