Efficient Single Correspondence Voting for Point Cloud Registration.

Area of Science:

• Computer Science
• Robotics
• Geomatics

Background:

• 3D point cloud registration is vital for applications like autonomous driving and virtual reality.
• Challenges include noise, non-uniformity, partial overlap, and repeated local features in large scenes.
• Existing methods struggle with the scale and complexity of real-world point cloud data.

Purpose of the Study:

• To propose an efficient single correspondence voting method for large-scale 3D point cloud registration.
• To improve registration accuracy and robustness in challenging environments.
• To provide a novel algorithm addressing limitations of current registration techniques.

Main Methods:

• Introduced SCVC (Single Correspondence Voting) for efficient hypothetical transformation prediction (5 DoF) using one correspondence.
• Employed Hough voting to determine the final degree of freedom for transformation.
• Developed VDIR (Verification via Discrepancy in Registration) for robust transformation verification and optimal result selection.

Main Results:

• The proposed method achieves state-of-the-art performance on benchmark datasets (3DMatch, 3DLoMatch, KITTI, WHU-TLS).
• Demonstrated significant improvements in registration accuracy for both indoor and outdoor large-scale scenes.
• The algorithm effectively handles noise, non-uniformity, and partial overlap issues.

Conclusions:

• The novel single correspondence voting method offers an efficient and accurate solution for large-scale 3D point cloud registration.
• SCVC and VDIR provide a robust framework for addressing complex registration challenges.
• The method's state-of-the-art performance validates its effectiveness in diverse real-world applications.