International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
Figure 1. Geocoded PS set in 3D coordinate system. The colour 
codes the height 
targets with its 3D position and its deformation time series is 
obtained. In Figure 1 a 3D view of the geocoded point cloud is 
shown. It is quite easy to recognise the major urban structures 
in this point cloud. An automatic reconstruction of buildings as 
done with airborne laserscanning data is however quite difficult 
due to the irregular point distribution and the limited 
positioning accuracy. For that reason additional map data is 
used, which makes an assignment of PS to buildings possible. 
2.2 Building Outlines 
The GIS data used throughout this research has been digitised 
manually using Google Earth™. Due to that its level of detail 
and accuracy is limited. However, we believe it to be sufficient 
for the purpose of determining the affiliation of PS to buildings. 
The GIS data is shown in Figure 2 overlaid to a Google Earth™ 
view of the scene. Internally the map data is represented as a set 
of polygons. 
Since the data has been digitised piecewise, its level of detail as 
well as its accuracy are not homogeneous. In general, interior 
courtyards are not contained unless they exceed a certain size. 
In order to estimate the needed volume per building a mean 
height has been acquired manually for every outline polygon 
leading to a prismatic building model. While this is a good 
approximation in simple cases, it gets inaccurate for more 
complicated buildings. The prismatic building models can be 
Figure 2. The used building outlines overlaid to a Google 
Earth™ map. 
      
seen in Figure 6. 
3. METHODOLOGY 
The compilation of a PS density map over buildings requires 
essentially two processing steps. First the PS point cloud and 
the GIS data have to be geometrically aligned. Secondly, the PS 
have to be assigned to the buildings. Both steps are conducted 
using the planimetric positions of the PS only, since no 3D GIS 
data like a 3D city model are available. 
3.1 Geometrical alignment 
The alignment of the datasets with respect to each other is 
necessary since both may be systematically displaced from their 
true position. This is especially the case for the PS point cloud 
due to the necessity to choose a reference PS at zero elevation. 
An inappropriate choice will results in a point cloud displaced 
in the sensors viewing direction (Gernhardt et al, 2011). Since 
the GIS data have been digitised using Google Earth™, it 
cannot be considered accurate. However, we are just interested 
in a relative alignment of both datasets. Thus, the PS point 
cloud is shifted in order to match the map data, which is 
advantageous since all results are referenced to Google Earth™ 
by doing so. The methodology employed to estimate the 
misalignment between both datasets is a simplified version of 
the popular Iterative Closest Point algorithm (ICP) (Besl, 
1992). The coordinate transformation is assumed to be a two- 
dimensional shift, which simplifies the ICP procedure 
considerably. 
Initially PS located on building roofs and other structures 
within the building outlines have to be filtered out as they 
cannot have corresponding points in the GIS data. This is done 
with a filter similar to the one used in (Gernhardt et al, 2011) to 
remove PS located at facades. It essentially checks the height 
variance of all PS in a local neighbourhood around the point 
under investigation. If the variance exceeds a threshold, the PS 
is tagged as a facade PS and kept for the ICP procedure. While 
this is quite effective to filter out PS on building roofs and on 
the ground, it certainly does not remove PS at vertical structures 
inside the outline of the building like facades bounding interior 
courtyards. Those structures are in general not included in the 
used GIS data, which may be problematic in some cases. 
After filtering the ICP procedure is performed. For every point 
of the PS set a shift vector to the closest point in the map data is 
determined. The distance between a point and a polygon edge 
(i.e. a line segment) is defined following (Besl, 1992) and 
sketched in Figure 3. If the PS is located in the area shaded in 
grey, the distance is measured along the plumb line (PS2 with 
distance d2). Otherwise the distance is measured to one of the 
polygon points (PSI and PS3). For every PS all edges in a 
certain neighbourhood are checked. Given the point 
correspondences a shift can be estimated which minimises the 
sum of the squared distances between both datasets. 
Let (x;y; denote the coordinates of the PS set and (Xj, Y)) 
denote the coordinates of the set of corresponding points in the 
map data. 
    
PS1 
m 
ME 
. polygon edge 
Figure 3. Definition of the distance between polygon edges 
and PS. 
80