Feature Consistent Point Cloud Registration in Building Information Modeling

Area of Science:

• Computer Vision
• Geometric Computing
• 3D Data Processing

Background:

• Point cloud registration is crucial for Building Information Modeling (BIM) applications like measuring and simulation.
• Existing methods struggle with generalization due to varying sampling environments and normal ambiguity at object boundaries.
• Current approaches often prioritize spatial transformation accuracy over feature matching, limiting performance.

Purpose of the Study:

• To develop a more robust and generalizable point cloud feature extraction method for BIM.
• To address the challenges of normal ambiguity and improve accuracy in point cloud registration.
• To enhance feature matching capabilities alongside spatial transformation accuracy.

Main Methods:

• Proposed a boundary-encouraging local frame reference called PyramidFeature (PMD).
• PMD integrates point-level, line-level, and mesh-level information for feature extraction.
• Introduced a hybrid feature extraction supervision method to improve consistency.

Main Results:

• The proposed PyramidNet (PMDNet) demonstrated superior performance in point cloud registration.
• PMDNet achieved high scalability and accuracy, even with significantly different training and testing datasets (ModelNet40 and BIM).
• The method effectively overcomes normal ambiguity at boundaries, leading to more accurate transformations.

Conclusions:

• PyramidNet (PMDNet) offers a significant advancement in point cloud registration for BIM.
• The PMD feature extraction method enhances generalization and robustness in diverse environments.
• This approach improves the reliability of 3D data processing in BIM applications.