Supervoxel-based multi-level region growing method for surface point cloud extraction in intersecting tunnels

Abstract

Mine intersecting tunnels are characterized by multi-directional connectivity, irregular surface morphology, and complex ancillary pipelines and facilities. Existing methods often fail to achieve complete extraction of tunnel surface point clouds. Therefore, this paper proposes a precise extraction method for intersecting tunnel surface point clouds based on supervoxel-based multi-level segmentation. First, supervoxels are generated using a boundary-constrained and cross-surface segmentation approach, which are then classified by linearity from saliency features and used as processing units. Next, a local adjacency graph with adaptive neighborhood selection and fitting-residual optimization is constructed. To handle complex intersections, a multi-factor hierarchical seed ranking integrates geometric features and adjacency consistency, ensuring reliable initialization. Furthermore, a multi-level region growing criterion is proposed for different linearity types to achieve robust extraction of major surface clusters. Finally, convex hull boundary overlap post-processing detects and refines misclassified supervoxels, improving completeness. Experiments on three datasets show strong performance under intersecting structures, uneven surfaces, and high noise: IoU consistently exceeds 98%, reaching up to 99.61%, while the surface extraction precision(SP) and surface extraction recall(SR) remain above 99%, significantly outperforming the comparison methods. These results fully validate the robustness and applicability of the proposed method for complex intersecting tunnel surface point cloud extraction tasks.