RandNet-LDE: An enhanced RandNet for semantic segmentation of operational tunnel point clouds
Z. Han, X.Y. Xie, G. Tang · 2026
Operational tunnel point clouds are dominated by lining points while facility objects are small and sparse, making semantic recognition difficult. This paper proposes RandNet-LDE, a lightweight enhancement to RandNet, and evaluates it through optimized training and validation on the Gaojiashan Tunnel dataset. The approach integrates long-tail reweighting, two-stage decoupled learning, and boundary-aware auxiliary supervision to mitigate class imbalance, improve boundary discrimination, and preserve fine details. Comparative experiments against PointNet and vanilla RandNet show significant gains in Overall Accuracy (OA) and mean Intersection over Union (mIoU), with particularly strong improvements for minority facility classes such as luminaires, fans, and cameras. Moreover, RandNet-LDE markedly reduces adhesion and misclassification at lining–facility boundaries, yielding higher overall accuracy and better detail fidelity. These results demonstrate the synergistic effectiveness of the three strategies in complex engineering scenarios and provide an efficient, transferable solution for high-precision semantic segmentation of operational tunnel point clouds.