Plane Segmentation Based on the Optimal-Vector-Field in LiDAR Point Clouds.

Area of Science:

• Computer Vision
• 3D Data Processing
• Computational Geometry

Background:

• Point cloud segmentation faces challenges in accurately detecting and splitting overlapping planar regions.
• Current methods rely on local information, leading to over- and under-segmentation issues.
• A global constraint is needed to improve the accuracy of plane segmentation in point clouds.

Purpose of the Study:

• To present a new pipeline for accurate plane segmentation in point clouds.
• To address the shortcomings of local optimization in existing point cloud segmentation methods.
• To improve the precision, recall, completeness, and correctness of plane detection.

Main Methods:

• A two-phase segmentation process is proposed, involving a local and a global phase.
• The local phase calculates connectivity scores using surface normal variations and an optimal-vector-field for plane intersection detection.
• The global phase smooths local cues using graph-cut inspired techniques for improved segmentation.

Main Results:

• Achieved 94.50% precision and 90.81% recall on mobile LiDAR data.
• Obtained an average accuracy of 75.4% on an open benchmark dataset.
• Outperformed state-of-the-art methods in terms of segmentation completeness and correctness.

Conclusions:

• The proposed pipeline effectively addresses the limitations of local optimization in point cloud plane segmentation.
• The method demonstrates superior performance compared to existing techniques on real-world and benchmark datasets.
• This approach offers a more robust solution for accurate plane detection and splitting in 3D point cloud data.