Geometry-constrained dynamic extraction of complex roadway ontology point clouds

• Radial cloth-simulation converts filtering into morphology-based recognition. • Polar-axis constraint captures tubular roadway topology in MLS point clouds. • Outperforms SOR/low-pass with 99.0 % precision and 92.3 % recall. • Robust across three roadway scenes with diverse facilities and clutter. • Provides cleaner inputs for digital twins and deformation monitoring. High-fidelity three-dimensional digital models are a prerequisite for developing intelligent mining systems and digital twins. However, accurately separating roadway structures from auxiliary facilities and noise in raw point clouds acquired by mobile laser scanning remains a critical bottleneck and major challenge in current data processing workflows. Traditional filtering algorithms, which rely on local statistical properties, struggle to handle clustered noise that exhibits local features similar to those of roadway structures, revealing their inherent limitations. To address this issue, we propose a geometry-constrained dynamic extraction method for complex roadway point clouds. The central innovation of the proposed approach is a paradigm shift from local statistical analysis to global geometric constraints. Specifically, we deconstruct and reformulate the classical Cloth Simulation Filter (CSF) by introducing a novel radial gravity model to replace the conventional vertical gravity. Within a polar coordinate system aligned using principal component analysis (PCA), this model drives virtual cloth to contract centripetally. This process establishes a strong global geometric prior that enables the algorithm to capture the tubular topology of the roadway, thereby elevating point cloud separation from a local filtering problem to a morphology-based structural recognition problem. Experimental evaluations confirmed the effectiveness of the proposed method. Compared with manually annotated ground truth, the proposed method achieved 99.0 % precision, 92.3 % recall, and 91.4 % overall quality, significantly outperforming traditional low-pass and statistical outlier removal (SOR) filtering approaches. Furthermore, successful validation across three roadway scenes with distinct physical characteristics and varying complexities (including air ducts, conveyor belts, supports, and monorails), combined with its integration with a divide-and-conquer segmentation strategy, demonstrates its robustness and adaptability to practical engineering applications. This study introduces a novel technical paradigm for point cloud processing of roadways and other tubular structures and provides a precise and reliable engineering solution that markedly improves the quality of 3D mine data. The proposed method establishes a solid foundation for subsequent mine digital twin development, surrounding rock stability monitoring, and intelligent decision making.