Automated Trace Extraction of Rock-Tunnel Excavation Face from 3D Point Clouds Based on Boundary Point Pre-Grouping

Since discontinuities are the key factors controlling the stability of rock mass, the accurate extraction of the geometric information of discontinuities, including discontinuity traces and orientations, is of great significance for conducting stability assessment of surrounding rock. However, existing automated methods of trace extraction based on three-dimensional (3D) point clouds rely on time-consuming triangulation, and the trace growth process is highly susceptible to noise interference, leading to directional deviations and false connections. This study proposes an improved algorithm for the automated extraction of discontinuity traces based on orientation grouping without triangulation. Principal Component Analysis (PCA) is employed to calculate the normal vectors of each point in the 3D point cloud. Then uniform basis vectors are constructed by discretizing the hemispherical space through the icosahedron subdivision method, which reduces the computational cost of orientation grouping. For rapid identification of the dominant orientations, the density peak criterion and the weighted fast silhouette index are used to non-iteratively identify the dominant orientation centers and the optimal number of groups. On this basis, the boundary points of the discontinuities are identified, followed by oriented contraction to obtain more representative feature points. After feature point pre-grouping, the growth of the trace segments is achieved through direction constraints. The algorithm is validated using measured tunnel point clouds and public dataset. The results show that the pre-grouping mechanism can effectively suppress local noise interference and enhance the robustness of trace extraction in complex scenarios, increasing precision, recall and F1-score by approximately 36%, 5%, and 15%. The number and average length of traces identified by the proposed algorithm are 95 and 0.28 m, while those identified manually are 86 and 0.32 m, which relative errors for the total number of traces and the average trace length are 10.4% and 12.5%. These errors do not affect the BQ surrounding rock classification. Compared with the existing method based on triangulation, the proposed method achieves approximately 39% higher recall and 21% higher F1-score, while precision remained comparable within 2%. The processing time is reduced by approximately 8.5% under the same computational environment.