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

Plane segmentation of real-world 3D point clouds captured by LiDAR or depth sensors remains challenging due to data sparsity, noise, and complex geometric configurations such as stepwise and intersecting non-coplanar structures. To address these issues inherent in sensor-acquired data, this paper proposes a geometry-aware plane segmentation method that leverages supervoxel boundary adjacency, normal coherence, and projection-line fitting constraints. Supervoxels were generated using the toward better boundary preserved supervoxel segmentation (TBBS) algorithm, and their natural adjacency relationships were constructed based on boundary points. Subsequently, the supervoxels were initially clustered according to their normal information. Finally, the projected point clouds of adjacent supervoxel were fitted with straight lines, and the fitting errors were calculated to optimize the clustering results. Experimental results demonstrate that this method performs excellently in handling stepwise non-coplanar structures, effectively segmenting planar regions with significant geometric features. It shows particular advantages in cases involving stepwise non-coplanar and intersecting planes. On benchmark datasets, the method achieves precision and recall rates of (97.7%, 94.4%, 91.2%, 80.4%, 92.3%) and (98.9%, 95.7%, 93.7%, 84.8%, 96.0%), respectively, highlighting its effectiveness and robustness for practical 3D sensing applications.