Fast and Robust Iterative Closest Point.

Area of Science:

• Computer Vision
• Robotics
• Computational Geometry

Background:

• Iterative Closest Point (ICP) is crucial for rigid registration of 3D point sets in robotics and 3D reconstruction.
• Traditional ICP suffers from slow convergence and sensitivity to outliers, noise, and partial overlaps.
• Existing robust methods like Sparse ICP trade speed for improved robustness.

Purpose of the Study:

• To develop a novel Iterative Closest Point (ICP) algorithm that achieves both fast convergence and robustness.
• To enhance the accuracy and computational efficiency of 3D point set registration.

Main Methods:

• The classical point-to-point ICP is reformulated as a majorization-minimization (MM) algorithm.
• Anderson acceleration is introduced to expedite the convergence of the MM algorithm.
• A robust error metric using Welsch's function is integrated and minimized via the accelerated MM approach.
• The robust formulation is extended to point-to-plane ICP using a similar strategy.

Main Results:

• The proposed method demonstrates similar or superior accuracy compared to Sparse ICP on challenging datasets with noise and partial overlaps.
• The new ICP variant is at least an order of magnitude faster than Sparse ICP.
• Registration accuracy is improved on benchmark datasets while maintaining competitive computational time.

Conclusions:

• The Anderson-accelerated MM approach with a robust error metric provides an efficient and accurate solution for 3D point set registration.
• This work offers a significant advancement over existing ICP methods, particularly for real-world applications with imperfect data.
• The developed robust ICP methods are competitive in both accuracy and speed, making them suitable for demanding applications.