Page 650 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

The utilized Ground Removing algorithm is summarized below:

Figure 7. Ground Removing algorithm

2.2. Building Extraction and Modeling
There are combined data from MLS and ALS point cloud. First of all, using ALS data to recognize building foot print.
Then, facade segmentation is implemented by utilization of MLS data. Finally, building utility modeling is obtained.

2.2.1 Building Foot Print Recognition: ALS point cloud is used to extract building. ALS data has more accurate
information about building roofs than MLS data, but MLS data doesn’t have any information about roofs. This is because of
the fact that ALS has both down-look view for recording point clouds and notably higher elevation than building roofs. It’s
true that MLS has high density of acquisition data, but includes information about other objects such as trees and cars that
decreases accuracy of propose algorithm for extracting building roofs. According to these reasons, MLS data is not used in
this step and just ALS data is applied. In this section, for extracting roofs, Digital Surface Model (DSM) is used.
An approach is used for building reconstruction from a single DSM. First, 2D supports of urban structures are extracted.
Second, a Gibbs model is utilized for placing for 3D blocks on 2D supports. Finally, for finding optimal shape of 3D blocks
a Bayesian decision is used with a RJMCMC sampler. This approach has very good results from a single DSM and works
more efficiency on different data resolutions [18].
Since DSM and point cloud are georeferenced, after specifying buildings on DSM, each pixel can be identified on point
cloud data. So, location of each building is recognized into point cloud.

2.2.2 Facade Segmentation: In the process of building extraction, just information about roofs are available. To extract
facades, MLS data is used. Boundary of each building is determined in area along X and Y axis. In ALS dataset, top surfaces
such as roofs have dense points, but vertical surfaces such as walls usually have sparse points or even no points. In top view,
vertical walls are exactly under the roof, also MLS collects density data from walls. Due to ALS and MLS data are
georeferenced, derived facades’ information from MLS data is possible.
For extracting building boundary, first building points are identified. Second, to avoid losing information split-and-merge
segmentation based on octree is utilized. Then, height of points is applied to detect boundary points. Finally, Hough transform
is applied to fit straight lines in a point dataset [19].
A very significant challenge in collecting data is that Mobile Laser Scanner not only extracts the walls but also derives
a small part of the roofs of each floor of the building. Passing laser beam through the windows makes to collect information
of piece of roof. An elevation feature is used to eliminate extracted part of roof. By selecting elevation method, k neighbor
points around each single point are chosen. After that, minimum and maximum elevations are identified among selected
points. Differences between minimum and maximum elevation and all of the calculated values for difference are transferred
to the range of [0,1].
Points whose calculated difference is close to zero are considered as roof points and removed from collected data with a
certain threshold. Then, all points of walls are remained.
For facade segmentation, Ransac algorithm is used. First of all, in Ransac algorithm, a point is selected randomly. Second,
k-nearest neighbor (k-NN) points are chosen around each point. Third, a line model is fitted to points. So, defined the inlier
and outlier points. In addition, all of steps for every inlier point are implemented and refit a model as long as number of inlier
points not to be increased. Finally, boundary of each wall is determined and points of wall are defined from MLS data. Ransac
algorithm is represented in Figure 8.

5