A Few-Shot Learning-Based Point Cloud Semantic Segmentation Network for Tunnel Lining Inspection

Abstract

Next-generation 6G networks will significantly advance the development of integrated sensing, communication, and computing (ISSC) systems, particularly in collection and processing of point cloud data. High bandwidth and low latency offered by 6G enable sensors to generate high-resolution point cloud data more efficiently, providing precise geometric information for tunnel lining inspections. As a key application within ISSC systems, tunnel lining detection has garnered widespread attention in the transportation and infrastructure sectors, helping to enhance the structural stability of tunnels and ensure their long-term safe operation. However, current tunnel inspection methods often require extensive experimental data and struggle to effectively extract features from tunnel objects. In this article, we propose a novel point cloud semantic segmentation (PCSS) network built upon few-shot learning for tunnel detection, capable of segmenting various essential elements within the tunnel, such as bolts, pipes, and tracks. First, due to the prevalent issue of sample imbalance in tunnel point cloud data, we introduce few-shot learning to tackle this challenge, enabling the model to perform effective semantic segmentation with limited data samples. Second, recognizing that different objects and structures within the tunnel scene may exhibit significant scale variations, we employ multiembedding networks to capture features at various scales within the point cloud data. Additionally, we propose a heterogeneous feature interaction (HFI) module to merge features derived from distinct embedding networks.