Curvature-Aware Point-Pair Signatures for Robust Unbalanced Point Cloud Registration.

Area of Science:

• Computer Vision and Geometry Processing
• 3D Point Cloud Registration

Background:

• Existing point cloud registration methods struggle with unbalanced pairs (varying spatial extent and point density).
• Accurate local registration is difficult due to scale variations and uneven density distributions in unbalanced point clouds.

Purpose of the Study:

• To develop a novel registration method for unbalanced point cloud pairs.
• To address the challenges of scale variation and uneven density in point cloud registration.

Main Methods:

• Keypoint detection in both source and target point clouds.
• Establishing initial one-to-many correspondences using local feature descriptors.
• Employing a novel Local Point Cluster Structure Feature for outlier rejection.
• Transformation hypothesis generation and evaluation using RANSAC.

Main Results:

• The proposed method significantly outperforms state-of-the-art methods on the KITTI-UPP benchmark in registration success rate and computational efficiency.
• Achieved competitive results on the real-world TIESY Dataset, demonstrating applicability and generalizability.

Conclusions:

• The Local Point Cluster Structure Feature is effective for outlier rejection in unbalanced point cloud registration.
• The proposed method offers a robust and efficient solution for registering point clouds with significant disparities.