关键词: 点云, 巷道, 级联式BP神经网络, 特征提取, 分类

Abstract: Efficient classification of tunnel point cloud data is crucial for safety monitoring and 3D reconstruction in underground transportation and mining operations, as it facilitates the comprehensive exploration and utilization of point cloud data. This study addresses issues in existing tunnel point cloud classification methods, such as noise sensitivity, low processing efficiency, and susceptibility to overfitting, by proposing a cascaded backpropagation (CBP) neural network classification method optimized with an early stopping mechanism and adaptive parameter tuning. Firstly, the Trimble RealWorks software is used to separate tunnel and ground point clouds. Then, local geometric features are extracted using spherical neighborhood space and covariance matrix eigenvalues to construct feature vectors. Finally, an improved CBP network is employed to hierarchically classify internal tunnel lighting equipment, signage, and various pipelines, thereby enhancing classification efficiency and accuracy. Experimental results demonstrate that the improved CBP neural network achieves high accuracy and reliability in tunnel point cloud classification, significantly improving data processing efficiency and providing data support for tunnel maintenance, renovation, and safety management.

Key words: point cloud, roadway, cascaded BP neural network, feature extraction, classification