Sparse-to-Dense Matching Network for Large-Scale LiDAR Point Cloud Registration.

Area of Science:

• 3D Computer Vision
• Robotics
• Geospatial Data Processing

Background:

• Point cloud registration is crucial for 3D scene understanding and analysis.
• Existing learning-based methods for LiDAR point cloud registration face challenges with large-scale outdoor data, including computational cost (dense-to-dense) and accuracy (sparse-to-sparse).

Purpose of the Study:

• To propose a novel Sparse-to-Dense Matching Network (SDMNet) for efficient and accurate registration of large-scale outdoor LiDAR point clouds.
• To address the limitations of existing dense-to-dense and sparse-to-sparse registration methods.

Main Methods:

• SDMNet employs a two-stage registration process: a sparse matching stage followed by a local-dense matching stage.
• The sparse matching stage utilizes a spatial consistency enhanced soft matching network with outlier rejection and a neighborhood matching module.
• The local-dense matching stage refines correspondences by matching points within local spatial neighborhoods of high-confidence sparse matches.

Main Results:

• SDMNet achieves state-of-the-art performance on three large-scale outdoor LiDAR point cloud datasets.
• The proposed method demonstrates high efficiency in registering large-scale outdoor point clouds.
• The combination of sparse and local-dense matching significantly improves registration accuracy and robustness.

Conclusions:

• SDMNet offers an effective solution for large-scale outdoor LiDAR point cloud registration.
• The novel sparse-to-dense approach balances efficiency and accuracy, outperforming previous methods.
• This work advances the field of 3D computer vision for real-world applications.