Inferring High Resolution Terrain, Vegetation, and Lines of Sight Models from Point Cloud Data
thesisposted on 18.10.2016 by Benedetto Vitale
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
With the diffusion of technologies which enables the acquisition of the 3-Dimensional (3D) structure of different surfaces, there has been a great focus on the information brought and the uses that are possible with this kind of data. One of the field in which this technology has been widely used is ecology, given that 3D data offers the possibility of analyzing trees and their characteristics directly from the point cloud structure by means of techniques aimed at isolating ground points from non-ground ones, procedures with the goal of segmenting individual trees in the cloud and analysis with respect to the trees point structure obtained in order to derive the trees properties. In this thesis we proposed a framework which infers information on vegetation and terrain starting from raw point cloud data of forested environments, classifying first the points in the point cloud as ground and non-ground by means of a multiscale curvature algorithm proposed in the literature, then segmenting individual trees among the vegetation points by means of a new segmentation algorithm proposed. By leveraging on the information extracted for terrain and vegetation, the framework infers also the visual obstruction potential of the identified vegetation elements by means of a supervised inference approach. The first phase of the supervised inference consists in extracting a suitable set of features from the vegetation points structures, in which we tried to extract features both through a manual approach, meaning that we chose a specific set of features to extract, and through neural networks in an automatic way. In the second phase instead, different machine learning algorithms were used in order to infer the visual obstruction potential of each vegetation element using the features extracted in the previous phase. We also provided a practical example among the possible uses of the information extracted by the framework, which consists in exploiting the framework results to analyze lines of sight among the individuals located in the environment. Another possible use of the information extracted by the framework could providing support in generating an immersive experience in forested environments, experiencing the habitat exactly as animals living in it. The reason behind the generation of an immersive environment is that interactions among animals of the same or different specie significantly influence their behavior under many valuable aspects, thus being able to track and analyze their movements and actions in a detailed and immersive way would bring useful insight to animal behavior studies.