Labelling Of Objects From Point Clouds For Building Recognition In Suburban Scenes

This thesis is to investigate the applicability of frequently used 3D point cloud classification approaches and explore as well as optimize the most influential parameter,for the sake of extracting building objects in large-scaled suburban scenes from different data sources(i.e.Terrestrial Laser Scanning(TLS)and photogrammetric point clouds).In this work,Experiments are conducted with variant supervised classification-based strategies,including several commonly used and well-performed feature extraction methods and classifiers.Furthermore,3D point datasets produced by different measurements are used in experiments to evaluate if the well-performed methods for labelling building points of 3D point clouds on TLS point cloud with less percentage of noise can perform well in a more complex scene,the photogrammetric point cloud with higher noise level and more outliers.Besides,a parameter optimization experiment on both feature extractors and supervised classifiers is conducted to figure out which parameter plays the most important role of building points labelling process.All these experiments is designed in order to provide some experimental evidence for establishing a well-designed 3D point cloud classification approach for labelling the complex real-world and some empirical knowledge for parameter selection.The algorithm and experiments conducted in this work are implemented with the help of functionalities from Point Cloud Library(PCL).