Geometric Primitives in MLS Point Clouds Processing
Date
2020-04-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Mobile Light Detection and Ranging (LiDAR), as an active remote sensing system, has become an accessible street-level mapping technology in the last decade due to its ability to collect accurate and dense 3D point clouds efficiently. Although tremendous effort has been made to LiDAR data processing, there still exist many problems in everyday tasks ( e.g., segmentation and detection). In this thesis, the LiDAR data processing is re-visited from a geometric-primitive perspective, with the hope that existing problems can be partly solved or even well addressed by tapping the potential of geometric primitives. A survey on geometric primitive extraction, regularization and their applications is presented for the first time. In this review, geometric primitives that consist of a group of discrete points are categorized into two classes: shape primitives (e.g., planes) and structure primitives (e.g., edges). The rest of this thesis focuses on geometric primitives in mobile LiDAR data processing. A fast 3D edge extraction method which consists of finding and linking edge candidates is proposed and tested in large-scale scenes. Given extracted edge clusters, a new facade separation method for mobile LiDAR point clouds is developed, based on which connected facades are separated into facade instances for the first time. To explore the potential of plane primitives in mobile LiDAR data processing, a novel instance-level building detection method based on plane primitives extracted from original point clouds is proposed. After that, a new point cloud segmentation algorithm that succeeds in separating buildings and vegetations is presented. The main contribution lies in using plane priors to improve segmentation accuracy. For line primitives, a new extraction method is presented in this thesis, which can extract multiple primitives simultaneously from projected point clouds. Based on extracted line segments, a graph-based method is presented to construct 2D building footprints. Last but not least, this thesis also introduces the energy-based ``hypothesis and selection" (HS) framework to object detection and segmentation in LiDAR point clouds for the first time. The adapted frameworks are proved to be flexible and effective according to extensive experiments in different applications.
Description
Keywords
LiDAR, Building instances, Point clouds, Geometric primitives, Segmentation
Citation
Xia, S. (2020). Geometric Primitives in MLS Point Clouds Processing (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.