Training-free Semantic Segmentation of Shield Tunnel Point Clouds

Abstract. Semantic segmentation of shield tunnel point clouds provides valuable information for checking assembly quality, deformation, or defects. To achieve efficient semantic segmentation of shield tunnel point clouds, Tunnel-NN, a training-free non-parametric network is proposed. Tunnel-NN is evaluated using a dataset of five point clouds of shield tunnel rings and compared to two common baselines, namely PointNet and PointNet++, demonstrating comparable performance. To further enhance its performance in large-scale point cloud applications, a sector splitter sampling method is introduced based on tunnel section geometry. This approach preserves local geometric features while reducing the size of input data. Test results indicate that the sector splitter significantly improves the segmentation accuracy of Tunnel-NN and also benefits PointNet and PointNet++. Compared to trainable deep learning algorithms, Tunnel-NN achieves similar performance on datasets with few training examples without the need for training, highlighting its potential for broad engineering applications.