Robust and Fast Point Cloud Registration for Robot Localization Based on DBSCAN Clustering and Adaptive Segmentation.

Area of Science:

• Computer Vision
• Geometric Computing
• 3D Data Processing

Background:

• Traditional Normal Distribution Transform (NDT) algorithms struggle with initialization, leading to local feature loss and mapping errors.
• Point cloud registration is crucial for 3D data analysis but often limited by scope and efficiency.

Purpose of the Study:

• To propose a novel point cloud registration approach, Clustering and Segmentation Normal Distribution Transform (CSNDT), to enhance registration scope and efficiency.
• To address the limitations of traditional NDT algorithms in handling local features and initialization.

Main Methods:

• Developed CSNDT by integrating an adaptive cell partitioning strategy into point cloud registration.
• Incorporated a judgment mechanism based on standard deviation and correlation coefficient into DBSCAN for improved adaptive clustering.
• Implemented straight-line point cloud clustering with adaptive grid cells and cell segmentation based on cell lengths.

Main Results:

• The proposed CSNDT algorithm demonstrates superior robustness and precision in point cloud registration.
• CSNDT offers improved matching efficiency compared to Iterative Closest Point (ICP) and standard NDT algorithms.
• Adaptive cell partitioning and segmentation effectively extend registration range and enhance accuracy.

Conclusions:

• CSNDT provides a more robust and efficient solution for point cloud registration challenges.
• The adaptive clustering and segmentation techniques significantly improve the performance of NDT-based registration.
• This method offers a promising advancement for applications requiring accurate 3D point cloud alignment.