Adaptively unsupervised seepage detection in tunnels from 3D point clouds

Abstract

To address the low efficiency and subjectiveness in manual visual inspection in operational tunnels, a hybrid approach that integrates the 3D Otsu method and K-nearest neighbors are developed for seepage detection from point clouds. Besides, the Delaunay boundary distance threshold method is used to adaptively extract the tunnel's inner surface. The 3D Otsu method preliminarily segments the metro tunnel point cloud between seepage and non-seepage parts according to the RGB value. The KNN method is followed to denoise the misidentified seepage points by the 3D Otsu method. A real tunnel case in the Wuhan metro system, China, is used to demonstrate the applicability and efficiency of the developed approach. Results indicate that the proposed approach is effective and stable for noise reduction, where 33.26% of the point clouds are identified as noise data and are later eliminated. Compared to the 3D Otsu method and the KNN method alone, the integrated 3D Otsu-KNN method can achieve a higher detection accuracy. The proposed approach can also achieve a comparable detection accuracy to the supervised learning method (e.g. support vector machine). It can be used as a decision tool to automate seepage detection and quantify the severity of seepage in underground tunnels.