Robust 3D point cloud registration based on bidirectional Maximum Correntropy Criterion.

Area of Science:

• Computer Vision
• Robotics
• Geometric Processing

Background:

• 3D point cloud registration is crucial for tasks like 3D reconstruction and object recognition.
• Traditional methods often struggle with real-world data containing significant noise and outliers.
• Mean Square Error (MSE)-based methods are sensitive to non-Gaussian noise, limiting their robustness.

Purpose of the Study:

• To develop a robust 3D point cloud registration algorithm.
• To address the limitations of MSE-based methods in handling noisy and outlier-corrupted data.
• To improve the accuracy and reliability of point cloud registration using information theory.

Main Methods:

• Utilized the bidirectional Maximum Correntropy Criterion (MCC) for robust point correspondence.
• Applied MCC as a probability measure to penalize noisy points during registration.
• Integrated fixed-point optimization with the iterative closest point (ICP) algorithm for parameter estimation.
• Employed bidirectional measures to maximize overlapping regions and prevent local minima entrapment.

Main Results:

• The proposed MCC-based algorithm demonstrated superior performance in handling non-Gaussian noise and large outliers.
• Bidirectional measures effectively avoided registration results being trapped in local minima.
• Experimental comparisons confirmed the algorithm's robustness and good performance under noisy conditions.

Conclusions:

• The bidirectional MCC approach offers a more robust solution for 3D point cloud registration compared to MSE-based methods.
• The algorithm effectively leverages information theory for enhanced point cloud processing.
• The proposed method provides a reliable tool for applications requiring accurate 3D registration in challenging environments.