Registration of 3D point clouds using a local descriptor based on grid point normal.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Processing

Background:

• Coarse-to-fine methods are crucial for point cloud registration in 3D reconstruction.
• Partial overlap in point clouds presents challenges for accurate coarse registration.
• Existing methods often struggle with accuracy when point clouds have limited overlap.

Purpose of the Study:

• To propose a novel 3D local feature descriptor, GNDAS, for accurate coarse registration of partially overlapping point clouds.
• To enhance the accuracy and robustness of 3D point cloud registration.
• To improve the efficiency of the overall registration process.

Main Methods:

• A new descriptor, grid normals deviation angles statistics (GNDAS), is proposed.
• GNDAS divides local surfaces into grids and analyzes normal deviation angles.
• Descriptor matching and transformation estimation generate initial registration, followed by ICP refinement.

Main Results:

• The GNDAS descriptor demonstrates high descriptiveness and robustness to low-level noise and varying mesh resolutions.
• Experimental results show superior accuracy and efficiency compared to existing methods.
• The proposed coarse-to-fine registration approach significantly improves results on public datasets.

Conclusions:

• The GNDAS descriptor offers a robust and descriptive solution for coarse point cloud registration.
• The proposed registration method achieves state-of-the-art accuracy and efficiency.
• This work advances 3D reconstruction techniques, particularly for challenging partially overlapping point clouds.