Diploma and Master Theses (authored and supervised):
"Integration of segmentation in point cloud classification";
Supervisor: N. Pfeifer, M. Pöchtrager;
Department für Geodäsie und Geoinformation,
final examination: 2019-09-05.
High density point clouds of airborne laser scanning measurements can be used to identify several object classes as buildings, vegetation or water surface. Point cloud segmentation can support classification and further feature extraction considering that segments are logical groups of points belonging to the same object class. This thesis presents a segment-based classification method, in which depending on the point features and attributes in the pre-processing step, a segmentation algorithm is executed seven times on the point cloud in order to extract point clusters belonging to different classes. These parts are used as a training set for classifying vegetation, building roof and wall, ground, vehicle, water and power lines. It is aimed that these segments as incorporated in the classification process will provide results comparable with the state of the art. Airborne Laser Scanning (ALS) point cloud datasets from the Municipality of Vienna are used, in an area where all the classes of interest are present. The whole workflow was implemented in OPALS, a modular software developed from TU Wien with the purpose to process airborne laser scanning data. The developed approach shows high classification results (overall accuracy is 86.9 %) based on the manual classified dataset and an internal accuracy of 94.8 % of the training data. However, there are some limitations of the method in terms of robustness and universal applicability, which were detected when applying the trained model to the test data. This expresses the importance of calculating accurate point attributes and performing a proper segmentation and classification of the objects of interest.
Electronic version of the publication:
Created from the Publication Database of the Vienna University of Technology.