Practical assessment of masonry tunnel joint segmentation using topological machine learning

Abstract Condition assessment of masonry lined tunnels is time consuming and labor intensive. Recently developed digital workflows enable structural models to be created automatically, reducing analysis time. As part of these procedures, it is important to be able to identify the location of each masonry block. Masonry joints can be segmented by applying deep learning to 3D point clouds obtained by lidar. However, these models often fail to separate block instances, reducing the effectiveness of subsequent analysis. Recent developments in topological loss functions enable models to more accurately connect detected structures. While these can be applied to better isolate individual masonry blocks, their performance depends on the selected training data, and so further investigation is required to enable the method to be applied effectively to different structures. This study investigates the ability of topological loss functions to enable deep learning models to operate on different tunnels with varying lining properties. By focusing on possible workflows for real world application of these methods, the study shows how training data type and origin impact performance. Block instance segmentation performance is evaluated directly using a new Blockwise Intersection Over Union metric. With this metric, training data volume and variety is shown to be a bigger driver of segmentation performance than either similarity between training and testing datasets or choice of loss function.