关键词: 隧道点云　, 语义分割　, 动态平衡　, 多尺度融合　, 注意力机制

Abstract: Existing methods suffer from insufficient accuracy when handling small target categories such as tunnel lining cracks and seepage,and struggle to effectively extract spatial feature information across different scales.This paper proposes a dynamic balancing and multi-scale fusion-based semantic segmentation method for tunnel point clouds,aiming to address the is sues of imbalanced class distribution and multi-scale geometric variations in tunnel point cloud data.Firstly,a multi-scale fea ture fusion module is introduced,which extracts local-to-global features with varying receptive fields through parallel convolu tional branches and adaptively fuses multi-scale features using an attention mechanism.Secondly,a dynamic balancing loss function is designed to adaptively adjust class weights based on sample counts and segmentation difficulty,thereby alleviating class imbalance.Tests on a self-constructed tunnel point cloud dataset show that,compared with various semantic segmentation methods such as PointNet++,CNN,GNN,and Transformer,the proposed method achieves a maximum improvement of 4.3 per centage points in the Intersection over Union (IoU) metric,reaching 96.8%.For critical small-target categories such as cracks and water seepage,the IoU is improved by 6.2 and 9.8 percentage points,reaching 98.5% and 95.3%,respectively.For defect targets with an area smaller than 0.1 m2,the detection accuracy reaches 98%.Correspondingly,the F1-score for cracks and wa ter seepage are 0.92 and 0.89,respectively,which are significantly better than those of other methods.This method provides a new technical solution for the intelligent detection of structural defects in tunnels,and is of great significance for ensuring the safety of tunnel operations.