|M.Sc Student||Tomer Shem-Tov|
|Subject||Parallel Computational Methods for Laser Scanning Derived|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Filin Sagi|
|Full Thesis text|
Technological advances in recent years are leading to an increase in the volume of geospatial data that are acquired. Along with that trend, affordable parallel computing hardware is spreading rapidly and becoming a major part of the computing spectrum. One prominent hardware component is the graphic processor unit (GPU) which is installed on the graphics card. The GPU is a powerful parallel processor that can act not only as a graphical processor, but also as a general-purpose one. Moreover, as low-end graphic cards equipped with such GPU can now be found on almost any computer, parallel computing can now be performed on almost any machine. As a result, GPUs can be regarded as an affordable means to accelerate common computations.
Following this trend, the study of parallelization of existing algorithms becomes an active research field and many studies of the application of GPU-based processing have been reported. Nonetheless, application of GPU-based computing to geospatial data processing is limited, especially concerning processing of laser scanning data. This limited application can be attributed to the recent introduction of this technology, the irregular 3D point distribution, which complicates its processing, and the large volume that characterizes such sets. The advantages of using such technology for such data are however clear, as the potential speedup of time-consuming processes can shorten the overall processing time.
This thesis studies the application of GPU-based computing to laser scanning data-related algorithms. It focuses on the application of parallel computing to filtering algorithms, which aim to generate digital terrain models. The thesis studies both local and global filters and concerns data management and geospatial-related algorithms as well as mathematical operations. Additionally, it studies the theoretical and practical limitations that should be taken into consideration when paralleling a program, including speedup limits and hardware implementation-dependent considerations. Application of the developed methods shows that an order of magnitude level of improvement can be reached. Thereby, they demonstrate that parallel methods have an important contribution for efficient processing of laser scanning derived data.