Clique-like Point Cloud Registration: A Flexible Sampling Registration Method Based on Clique-like for...

Area of Science:

• Computer Vision
• Robotics
• 3D Perception

Background:

• 3D point cloud registration is crucial for sensor-based 3D perception.
• Traditional methods like RANSAC struggle with low overlap, outliers, and computational cost.

Purpose of the Study:

• To develop a novel 3D registration method, CL-PCR, to overcome limitations of existing techniques.
• To improve robustness against low overlap and outliers in point cloud registration.

Main Methods:

• Introduced CL-PCR, a novel method based on maximal cliques and the SC^2-PCR framework.
• Constructed a graph matrix for correspondence compatibility and used clique-like subsets for consensus.
• Employed SVD for transformation hypothesis computation and selected the best based on evaluation metrics.

Main Results:

• CL-PCR demonstrated superior registration performance, particularly Fast-CL-PCRv1, on 3DMatch/3DLoMatch datasets.
• The method showed enhanced robustness against low overlap and reduced outlier influence.
• Validated effectiveness and practicality with real-world data.

Conclusions:

• CL-PCR offers a robust and accurate solution for 3D point cloud registration.
• The method's ability to leverage smaller sampling subsets improves performance in challenging environments.
• CL-PCR represents a significant advancement in 3D perception for sensors.