Segmentation of Mobile LiDAR Point Clouds in Urban Environment
Date
2018-09-19
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Mobile LiDAR System (MLS) is a popular tool which collects 3D information while vehicles drive along urban streets. Nowadays, MLS becomes more and more important in obtaining 3D point clouds because of its efficiency and cost-effectiveness. One demanding issue is how to segment objects from MLS point clouds accurately and efficiently. Due to the fact that LiDAR point clouds are noisy, uneven and the lack of topology, the object segmentation has become a challenging task. This thesis aims to provide promising solutions for segmentation of MLS point clouds. In the beginning, this thesis proposes an elevation-based method to split MLS point clouds into ground points and off-ground points. Then, it presents an accurate method to extract road curbs from ground points. Contributions of the proposed curb extraction method include that, it formulates an energy function to extract candidate curb points, and it completes curb paths by a new least cost path model. As this curb extraction method works on the 3D points directly rather than on the 2D projected data, there is no loss of 3D geometry information, which improves the extraction performance. For the off-ground object segmentation, two new methods are introduced in this thesis. The first method is an optimal hierarchical clustering (OHC) approach. The cluster combination in the hierarchical clustering is formulated as a problem of the bipartite graph matching, which is optimally solved by the minimum-cost perfect matching (McPM) of a point-based graph. The second method is a component-level segmentation approach for dealing with the separation of overlapping objects. Contributions of the second method include that, it formulates an optimal-vector-field to provide the component consistency information for object segmentation, and it presents a strategy for multi-object segmentation using binary labels only. Due to the fact that power lines are divided into pieces by the previous object segmentation methods, the final work is to extract power lines from input data. Contributions include that, it proposes an estimator to enhance the segmentation robustness against Gaussian noise, and it groups segmented components into individual power line spans based on linear structure information.
Description
Keywords
Mobile Mapping, Laser Scanning, Point Clouds, Computer Vision, Vegetation Mapping
Citation
Sheng, Xu. (2018). Segmentation of Mobile LiDAR Point Clouds in Urban Environment (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/33050