Plane Segmentation in Sensor-Acquired 3D Point Clouds Using Supervoxel-Based Geometric Constraints.

Area of Science:

• Computer Vision
• 3D Data Processing
• Geometric Modeling

Background:

• 3D point cloud segmentation is difficult due to sensor noise, data sparsity, and complex geometries like stepwise and intersecting planes.
• Existing methods struggle with the inherent challenges of real-world sensor-acquired data.

Purpose of the Study:

• To propose a novel geometry-aware plane segmentation method for 3D point clouds.
• To address limitations in segmenting complex, non-coplanar structures in sensor data.
• To enhance the robustness and accuracy of plane segmentation in practical 3D sensing.

Main Methods:

• Utilized supervoxels generated by the toward better boundary preserved supervoxel segmentation (TBBS) algorithm.
• Constructed supervoxel adjacency based on boundary points and initial clustering using normal information.
• Optimized clustering by fitting projected point clouds of adjacent supervoxels with straight lines and calculating fitting errors.

Main Results:

• The method excels at segmenting planar regions with significant geometric features, especially stepwise non-coplanar and intersecting planes.
• Achieved high precision (up to 97.7%) and recall (up to 98.9%) on benchmark datasets.
• Demonstrated excellent performance in handling data sparsity and noise inherent in sensor-acquired point clouds.

Conclusions:

• The proposed geometry-aware plane segmentation method is effective and robust for practical 3D sensing applications.
• It significantly improves the segmentation of complex geometric configurations in real-world 3D point clouds.
• The approach offers a reliable solution for applications requiring accurate plane extraction from LiDAR and depth sensor data.