3D features extraction from 3D point clouds

This work addresses an essential part of 3D perception and modeling for robotic applications. Autonomous robots must percept and represent their environments to reason what to do next to achieve given tasks. In the three dimensional space the perception procedures easily produce several thousands 3D points which are getting collected and transformed into 3D models. To access these huge models and to extract application dependent knowledge from them different approaches exist, but they are all facing the problem of finding inside the huge information sets that particular information which is needed. The problem addressed by this thesis is to identify features from environmental 3D point clouds to break down the search and problem spaces by segmenting the point cloud models. To do so this work provides a framework of algorithms for 3D feature extraction and segmentation. The methodology for the feature extraction and segmentation is based on an extensive state-of-the-art review which is included as a literature survey in this work. By side of filtering and feature extraction algorithm does the implemented framework include methods for segmentation which supports classification from structural objects like floors, walls and ceilings. In addition, a novel approach to detect structural shapes of an object by using the laser reflection intensity using difference-of-means strategy is also included and evaluated. The algorithms were evaluated using prior recorded environmental models gained by a rotational laser range under (3DLS-K2 developed at the Fraunhofer Institute for Intelligent Analysis and Information Systems).