Rock-Bolt Extraction from Tunnel Point Clouds Using Multi-Feature Learning and Spatial Consistency Refinement

Reliable inspection of rock bolts is essential for evaluating tunnel support quality in deep underground mining, yet manual assessment remains inefficient and subjective in complex field environments. Although three-dimensional point clouds provide a promising basis for automated inspection, robust rock-bolt extraction from real tunnel scenes remains difficult because bolt points are often sparse, weak, and locally confused with rough tunnel-surface structures. In this study, we propose a hierarchical framework for rock-bolt extraction from in-situ tunnel point clouds. The method combines local multi-feature representation, attention-based point-wise classification, spatial-consistency refinement, cross-block aggregation, and instance-level recovery. Experiments on real tunnel point clouds show strong point-level performance, with mean block-wise precision, recall, F1-score, IoU, and MCC of 0.854, 0.988, 0.910, 0.854, and 0.914, respectively. At the instance level, 120 matched bolt instances are recovered from 126 annotated bolts, yielding an instance precision of 0.992, an instance recall of 0.952, an instance F1-score of 0.972, and a mean matched IoU of 0.920. These results show that robust rock-bolt extraction from realistic tunnel point clouds is achievable when local feature representation, spatial consistency, and instance-level recovery are addressed jointly.