A deep learning approach for automation of joint sets recognition on 3D point clouds of rock mass surfaces APCOM 2021

In this paper, a methodology for a computerised recognition of joint sets on 3D point cloud models of rock masses is presented. The process starts with classifying joints on a 3D rock mass surface through training a deep network architecture and validated using manually labeled datasets. Then, individual joint surfaces are identified using the Density-Based Scan with Noise (DBSCAN) clustering algorithm. Subsequently, the orientations of the identified joint surfaces are computed by fitting least-square planes using the Random Sample Consensus (RANSAC). Finally, the joint planes are classified into different joint sets, and the dip direction and dip angle for each set are calculated. The performance of the proposed methodology has been evaluated using a case study. The results show that the presented procedure is fast, accurate, and robust.