A machine learning approach for the detection of supporting rock bolts from laser scan data in an underground mine

Abstract Rock bolts are a crucial part of underground infrastructure support; however, current methods to locate and record their positions manual, time consuming generally incomplete. This paper describes an effective method automatically supporting rock from 3D laser scanned point cloud. The proposed utilises machine learning classifier combined with descriptors based on neighbourhood properties classify all data points as either ‘bolt’ or ‘not-bolt’ before using the Density Based Spatial Clustering Applications Noise (DBSCAN) algorithm divide results into candidate bolt objects. centroids these objects then computed output simple georeferenced coordinates be used by surveyors, mine managers automated machines. Two classifiers were tested, random forest shallow neural network, network providing more accurate results. Alongside different classifiers, input feature types also examined, including eigenvalue geometric features popular in remote sensing community histogram common mobile robotics community. It was found that combination both sets provided strongest obtained precision recall scores 0.59 0.70 for individual 0.93 0.86 demonstrates model is robust noise misclassifications, still detected even if edge misclassified, there enough correct form cluster. In some cases, can detect which not visible human interpreter.