|Ph.D Student||Barnea Shahar|
|Subject||Terrestrial Laser Scans Processing Supported by Image Data|
|Department||Department of Civil and Environmental Engineering||Supervisors||Professor Sagi Filin|
|Dr. Victor Alchanatis|
|Full Thesis text|
Terrestrial laser scanners offer dense and accurate 3D data, thereby facilitating a detailed surface and objects description irrespective of their shape complexity. Consequently, they are rapidly becoming a standard technology for 3D modeling in variety of applications. Despite of their attractive output, the data volume and irregular distribution in 3D space turn the processing into a cumbersome task that requires considerable amount of involvement and lengthy interaction within the acquisition processes. All these steps necessitate a substantial amount of manually operated processes both in the field and in the office. They therefore call for increased level of automation in processing terrestrial laser scans.
This thesis studies automation of laser scanning related processes and focuses on two fundamental tasks, registration of individual scans into a common reference frame, and information extraction. Both are challenging due to the complexity and clutter of natural scenes, significant scale variations, occlusion, irregular point distribution, and noise. To facilitate both processes, the thesis proposes a data representation that enables computationally efficient processing and analysis. It also considers two additional sources which are recorded as part of the acquisition along with the geometric data. The first is the intensity data, which relates to the object radiometry, and the second is image data, acquired by a camera mounted on the scanner. The research explores the contribution of these two data sources to facilitate high-level of automation in processing point clouds. Experiments show registration accuracy reaching comparable results to a human operated one, while exhibiting robustness over wide baselines. In terms of information extraction, the proposed models yield physically meaningful segments, while in the extraction phase detection accuracy is of 99%.