DMS: Low-Overlap Registration of 3D Point Clouds With Double-Layer Multi-Scale Star-Graph

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Robotics

Background:

• 3D point cloud registration with low overlap is a significant challenge in computer vision.
• Existing methods struggle with identifying small overlapping regions and filtering correspondence outliers.

Purpose of the Study:

• To develop a robust method for 3D point cloud registration, particularly in scenarios with limited overlap.
• To improve the accuracy and reliability of correspondence detection and outlier rejection.

Main Methods:

• Proposed a Double-layer Multi-scale Star-graph (DMS) structure utilizing Multi-scale Neighbor Feature Star-graphs (MNFS) for initial correspondence candidate generation.
• Constructed Multi-scale Matching Star-graphs (MMS) and Multi-scale Correspondence Star-graphs (MCS) to identify mutual correspondences and filter outliers based on neighborhood consistency.
• Employed edge and vertex weighting criteria within MCS for robust correspondence selection.

Main Results:

• The DMS method successfully establishes initial correspondences using multi-scale neighborhood topology and feature similarity.
• Robust overlapping regions are detected by leveraging consistent neighborhood correspondences.
• Experimental results show superior robustness compared to state-of-the-art registration algorithms.

Conclusions:

• The proposed DMS structure effectively addresses the challenges of 3D point cloud registration with low overlap.
• The method demonstrates enhanced robustness in identifying correspondences and rejecting outliers.
• The developed technique offers a significant advancement for 3D computer vision applications requiring accurate point cloud alignment.