RGB-D Point Cloud Registration Based on Salient Object Detection.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Aligning red green blue-depth (RGB-D) point clouds is crucial for 3D reconstruction.
• Existing iterative closest point (ICP) algorithms struggle with noisy, partially overlapping, and unevenly distributed data.

Purpose of the Study:

• To develop a robust algorithm for aligning rigid, noisy, and partially overlapping RGB-D point clouds.
• To enhance the ICP algorithm's performance in challenging real-world scenarios.

Main Methods:

• Incorporated salient object detection (SOD) to focus on structurally significant foreground points.
• Introduced bidirectional color distance (BCD) for precise correspondence using color guidance.
• Integrated maximum correntropy criterion (MCC) and a trimmed strategy to mitigate noise and outliers.

Main Results:

• The proposed algorithm demonstrates superior robustness compared to existing methods on simulated and real-world data.
• Achieved satisfying 3-D reconstruction results for indoor scenes.
• Effectively handles data degradation and uneven point distribution.

Conclusions:

• The enhanced ICP algorithm provides a more robust solution for RGB-D point cloud registration.
• The combination of SOD, BCD, and MCC significantly improves alignment accuracy and 3D reconstruction quality.
• The method is effective in diverse and challenging environments.