GL-Net: semantic segmentation for point clouds of shield tunnel via global feature learning and local feature discriminative aggregation

Has gradually become the first choice of modern urban public transportation due to its advantages of safety and high-efficiency. Shield tunnel is an important type of subway tunnel, and its structural stability and safety play an important role in subway operation. The shield tunnels are prone to problems such as water leakage and tunnel collapse, which affect the safe operation of subways. Efficient monitoring methods are required to detect the status of subway tunnels. The data collection and accurate segmentation of key components of shield tunnels are the basis and key to the automatic monitoring of subway tunnels. This research presents a novel semantic segmentation method of three-dimensional (3-D) point clouds of typical structural elements (e.g., longitudinal joint, circumferential joints, bolt hole and grouting hole) in shield tunnel based on deep learning. In this method, we focus on how to make the network learn robust global features and complex local distribution patterns. Further, we propose a global and local feature encoding block (namely GL-block) to discriminatively aggregate local features while learning global representation. After multiple encodings by the GL-block, we design a global correlation modeling (GCM) module to establish a global awareness of each point. Finally, a weighted cross-entropy loss function is designed to solve the problem of unbalanced number of samples in each category of shield tunnel. In the experiments, we make a dataset of shield tunnel point clouds with a length of about 1,000 m collected by CNU-TS-1 (DU et al., 2018) mobile tunnel monitoring system, and use the dataset to train and test the segmentation ability of our method on the typical structural elements of shield tunnels. Experiments verify the effectiveness of our method by comparing with the other state-of-the-art 3-D point cloud semantic segmentation methods, and our method has an mIoU score of 73.02 %, which is at least 14.54 % higher than the other compared state-of-the-art networks. Also, we further verify the adaptability of our method to different tunnels and different laser scanning equipment, such as FARO, Leica and Z + F, and achieve very advanced performance.

• Autores: Li, Jincheng; Zhang, Zhenxin; Sun, Haili; Xie, Si; Zou, Jianjun; Ji, Changqi; Lu, Yue; Ren, Xiaoxu; Wang, Liuzhao
• Tipo de recurso: Article of investigation
• Fecha de publicación: 2023
• Institución: Corporación Universidad de la Costa
• Repositorio: REDICUC - Repositorio CUC
• Idioma: eng
• OAI Identifier: oai:repositorio.cuc.edu.co:11323/10613
• Acceso en línea: https://hdl.handle.net/11323/10613; https://repositorio.cuc.edu.co/
• Palabra clave: GL-Net; Semantic segmentation
• Rights: embargoedAccess
• License: Atribución-NoComercial-SinDerivadas 4.0 Internacional (CC BY-NC-ND 4.0)