Iterative most-likely point registration (IMLP): a robust algorithm for computing optimal shape alignment.

Area of Science:

• Computer Vision
• Computational Geometry
• 3D Data Processing

Background:

• Accurate rigid-body alignment of 3D shapes is crucial in various fields.
• Existing methods like Iterative Closest Point (ICP) can be sensitive to noise and outliers.
• Modeling anisotropic noise and local surface geometry remains a challenge for registration algorithms.

Purpose of the Study:

• To introduce a novel probabilistic registration algorithm for robust and accurate rigid-body alignment of 3D shapes.
• To enhance shape registration by effectively modeling various noise conditions and local surface characteristics.
• To improve upon the accuracy and robustness of existing point-cloud and mesh registration techniques.

Main Methods:

• Developed the Iterative Most-Likely Point (IMLP) algorithm, a probabilistic variant of ICP.
• Incorporated a generalized noise model into both correspondence and registration steps.
• Introduced a principal direction (PD)-tree search for efficient most-likely correspondence computation.
• Proposed a novel approach for solving the generalized total-least-squares (GTLS) sub-problem under a generalized noise model.

Main Results:

• IMLP demonstrates superior performance compared to ICP, GICP, CPD, GTLS-ICP, and A-ICP across diverse noise levels, outliers, and misalignments.
• The algorithm achieves high accuracy for both point-cloud and mesh representations of shapes.
• The proposed GTLS approach offers improved accuracy, efficiency, and stability over prior methods.

Conclusions:

• The IMLP algorithm provides a robust and accurate solution for rigid-body shape registration, outperforming existing state-of-the-art methods.
• The probabilistic framework effectively handles complex noise models and local surface properties, leading to enhanced registration quality.
• IMLP offers a practical and efficient advancement for 3D shape alignment tasks.