Generalized Extraction of Bolts, Mesh, and Rock in Tunnel Point Clouds: A Critical Comparison of Geometric Feature-Based Methods Using Random Forest and Neural Networks

Automatically identifying mine and tunnel infrastructure elements, such as rock bolts, from point cloud data improves deformation and quality control analyses and could ultimately contribute to improved safety on engineering projects. However, we hypothesize that existing methods are sensitive to small changes in object characteristics across datasets if trained insufficiently, and previous studies have only investigated single datasets. In this study, we present a cross-site training (generalization) investigation for a multi-class tunnel infrastructure classification task on terrestrial laser scanning data. In contrast to previous work, the novelty of this work is that the models are trained and tested across multiple datasets collected in different tunnels. We used two random forest (RF) implementations and one neural network (NN), as proposed in recent studies, on four datasets collected in different mines and tunnels in the US and Canada. We labeled points as belonging to one of four classes—rock, bolt, mesh, and other—and performed cross-site training experiments to evaluate accuracy differences between sites. In general, we found that the NN and RF models had similar performance to each other, and that same-site classification was generally successful, but cross-site performance was much lower and judged as not practically useful. Thus, our results indicate that standard geometric features are often insufficient for generalized classification of tunnel infrastructure, and these types of methods are most successful when applied to specific individual sites using interactive software for classification. Possible future research directions to improve generalized performance are discussed, including domain adaptation and deep learning methods.