Deep learning for point cloud segmentation with applications to sawmill industry

Abstract

Nowadays, the integration of digital technologies in enterprises is happening everywhere. The sawmill industry is no exception. The various stages of production at sawmills are being improved through the use of information technology and machine learning. In particular, the use of a 3D digital log model is very beneficial as it allows to simulate internal knot distribution and select optimal sawing pattern. One of the main problems is that 3D models of logs obtained using laser scanners often contain noise and various extraneous objects. This hinders the correct analysis. The noise can be filtered from the log using segmentation algorithms. Some solutions to this problem exist, but they require manual parameter configuration. The goal of this thesis was to solve this problem using modern deep learning approaches. More specifically, the aims were to review existing solutions for point cloud segmentation, select the most appropriate one, train and evaluate it. After consideration, RandLA-Net was selected because it can process large point clouds in a reasonable amount of time, and also due to its ability to accurately segment complex geometric structures. The selected method accurately distinguished noise and logs. It also demonstrated the ability to segment previously unseen data that was gathered in a different environment.