Support Vector Machines Based Filtering of Lidar Data (4063) |
| Mahmoud Gomah and John Trinder (Australia) |
Mr. Salah Gomah
Visiting Fellow
School of Surveying and Spatial Information System
The University of New South Wales
UNSW Kensington Campus - Sydney NSW 2052 Australia
School of Surveying and Spatial Information System
The University of New South Wales
Sydney
2052
Australia
|
| |
| Corresponding author Mr. Salah Gomah (email: m.gomah[at]unsw.edu.au, tel.: + 61 2 938 54197) |
| |
|
| [ abstract ] [ paper ] [ handouts ] |
|
Published on the web 2010-01-14
Received 2009-11-19 / Accepted 2010-01-14 |
| This paper is one of selection of papers published for the FIG Congress 2010 in Sydney, Australia and has undergone the FIG Peer Review Process. |
FIG Congress 2010
ISBN 978-87-90907-87-7 ISSN 2308-3441
http://www.fig.net/resources/proceedings/fig_proceedings/fig2010/index.htm
|
Abstract |
| Filtering is the process of separating on-terrain points (DTM) from non-terrain points. Filtering techniques can be classified into two main approaches: grid based filtering and raw data based filtering. For typically non-complex landscapes most of the existing algorithms do well, but the filtering of complex urban landscapes still poses the greatest challenges. In order to overcome this problem, we have suggested using image and lidar data fusion for filtering of these complex urban landscapes. This study introduces a method for filtering of lidar data based on Support Vector Machines (SVMs) classification.
First, the Digital Surface Model (DSM) was generated for the first and last pulses. Then the differences between the first and last pulses were computed. A total of 22 uncorrelated feature attributes have been generated from the aerial images, the lidar intensity image, DSM and the difference between the first and last pulses. Finally, Support Vector Machines (SVMs) were used to automatically classify buildings, trees, roads and ground from aerial images, lidar data and the generated attributes with the most accurate average classifications of 95% being achieved. Four kernel models (Gaussian Radius Basis Function (RBF); Linear; Polynomial; and Sigmoid) were tested for the filtering process. Further, we evaluated the contributions of the individual attributes to the quality of the filtering process
A binary image was then generated by converting the digital numbers of roads and grass to one while the digital numbers of buildings and trees were converted to zeros. Then all DSM’s pixels which correspond to a pixel value of one in the binary image were interpolated into a grid DTM. The quality of the derived DTM was improved by further reviewing the points that were classified as off-terrain points. To achieve this objective, a TIN was generated from the DTM and each off-terrain point was projected onto the TIN and its elevation calculated from the plane defined by the corresponding triangle. Points with heights within 30cm of the height calculated within the TIN were added to the on-terrain points and a new DTM generated. The process was repeated until no new points were added to the DTM.
To meet the objectives the filtered data was compared against reference data that was generated manually and both omission and commission errors were calculated. The results showed the relative importance of the difference between first and last pulses and its attributes for the filtering process.
|
| |
| Keywords: Laser scanning; Photogrammetry; aerial; LIDAR; fusion; classification; learning; filtering |
