Processing tree point clouds using Gaussian Mixture Models

Abstract

While traditionally used for surveying and photogrammetric fields, laser scanning is increasingly being used for a wider range of more general applications. In addition to the issues typically associated with processing point data, such applications raise a number of new complications, such as the complexity of the scenes scanned, along with the sheer volume of data. Consequently, automated procedures are required for processing, and analysing such data. This paper introduces a method for modelling multi-modal, geometrically complex objects in terrestrial laser scanning point data; specifically, the modelling of trees. The model method comprises a number of geometric features in conjunction with a multi-modal machine learning technique. The model can then be used for contextually dependent region growing through separating the tree into its component part at the point level. Subsequently object analysis can be performed, for example, performing volumetric analysis of a tree by removing points associated with leaves. The workflow for this process is as follows: isolate individual trees within the scanned scene, train a Gaussian mixture model (GMM), separate clusters within the mixture model according to exemplar points determined by the GMM, grow the structure of the tree, and then perform volumetric analysis on the structure.