Deep-learning-based longitudinal joint opening detection method for metro shield tunnel

Abstract

In this paper, a longitudinal joint opening detection method using a precise longitudinal segment joint extraction algorithm featuring deep neural networks (DNNs) is proposed. The proposed method consists of the following four steps. First, a mobile scanning system is employed to obtain three-dimensional metro shield tunnel point clouds. Then, two small DNNs, YOLOv5 and JLNet, were designed to accurately extract the longitudinal segment joint lines from the images generated from the scanned point clouds. YOLOv5 rapidly detects the approximate longitudinal segment joint areas, while JLNet precisely fits the joint lines. Subsequently, using the extracted segment joint lines, the points associated with different tunnel segments can be segmented accordingly. Finally, based on the tunnel segment point clouds, a joint opening angle calculation method that combines the cylinder projection and plane-fitting algorithms is proposed. Experimental results demonstrate that the proposed DNN-based method can accurately extract segment joint lines without being influenced by the tunnel equipment around the segment joints. The YOLOv5 network exhibited a classification accuracy of 0.9907 and a bounding box prediction error of 0.004. For the JLNet network, the line slope prediction error was 0.0072, with an intercept error of 1.53 pixels. The joint opening spatial distribution pattern was identified by comparing the joint opening angles in the deformed and undeformed tunnels. Additionally, the accuracy of the proposed method was evaluated, revealing that the joint opening angle detection external accuracy was 0.13°.