Groupwise rigid registration of wrist bones

Area of Science:

• Computer Vision
• Medical Imaging
• Geometric Processing

Background:

• Iterative Closest Point (ICP) algorithms are widely used for 3D shape registration.
• Existing ICP variants can introduce bias by requiring a reference shape or sequential registration.
• Accurate multi-shape alignment is crucial in fields like medical imaging and robotics.

Purpose of the Study:

• To develop an unbiased algorithm for aligning an arbitrary number of 3D shapes (N >= 2).
• To extend the symmetric ICP algorithm to handle multiple shapes without a reference.
• To improve registration accuracy and computational efficiency for multi-shape alignment.

Main Methods:

• An extension of the symmetric ICP algorithm was developed for unbiased alignment of N shapes.
• The method utilizes rigid transformations and scaling without a reference shape or registration order.
• A first-order approximation was proposed to estimate transformation parameters in closed form, reducing computational complexity.

Main Results:

• The proposed approximation method was shown to converge to the same solution as least-squares minimization.
• Experiments demonstrated smaller registration errors compared to reference-based or evolving mean shape algorithms.
• The algorithm achieved a computational complexity of O(N^2) and showed potential for parallelization.

Conclusions:

• The developed unbiased algorithm effectively aligns multiple 3D shapes with improved accuracy.
• The computationally efficient approximation enables the alignment of large numbers of shapes.
• This method offers a robust solution for multi-shape registration challenges in various applications.