there is no texture 
also not an option. 
detect the overall 
an extracting single 
ape is more robust 
ion by trees, cars, 
nto the scene. The 
ition for its overall 
CAD model of the 
according to the 
n how to select the 
Fritsch 2001). The 
ict the silhouette of 
on is then detected 
  
  
| for a given exte- 
low. 
nsform (GHT) to 
"Work for both the 
sional shapes in 
tect the shape no 
illy even scaled in 
sedom, since the 
litionally the GHT 
viation, which is 
1g is only a coarse 
in the image. 
rm (Hough, 1962) 
es or circles in an 
n of the Hough 
‘as a function of a 
on of that concept 
a simple analytic 
e edge magnitude 
ase, which has to 
so-called R-table 
s selected usually 
stance vector r of 
vith respect to its 
n the R-table. Of 
re can be several 
is performed, all 
1. For each edge 
y in the R-table, 
which possibly holds several, vectors ri. These vectors 
correspond to positions in the image, which receive a vote. The 
position in the image receiving the most votes at the end is 
selected as the position of the shape in the image. 
If the orientation of the object is allowed to vary, as is the case 
in our application, a separate R-table has to be computed for 
each discrete rotation angle. The same is true for scaling. Thus 
the formation of the R-tables in our case is quite complex and 
computationally expensive. 
P 
Figure 3: Properties of a single edge pixel P as recorded in the 
framework of the Generalized Hough Transform. 
For our implementation we made use of the HALCON image 
processing environment, which provides a shape detection 
mechanism based on the GHT (Ulrich, et. al. 2001). In order to 
compensate for the computational costs of large R-tables, this 
operator includes several modifications to the original GHT. 
For example it uses a hierarchical strategy generating image 
pyramids to reduce the size of the tables. By transferring 
approximation values to the next pyramid level the search space 
is drastically reduced. 
3. ORIENTATION 
The result of the GHT is a two-dimensional similarity 
transformation consisting of translation, rotation and scale. This 
transformation corrects for the misalignment of the shape to its 
actual appearance in the image. In other words for each three- 
dimensional point of the original CAD model rendered to the 
two dimensional image we can use the transformation to correct 
its location in the image. 
3.1 Spatial Resection 
This gives us the possibility to create point wise pseudo 
observations. We can select an arbitrary point from the CAD 
model of the building and project it onto the image using the 
calibration data of the camera and the initial approximated 
orientation. Then we use the two-dimensional transformation 
from the GHT to move the image point to its correct location. 
From a minimal configuration of three points we can compute 
an improved exterior orientation of the camera by 
photogrammetric spatial resection. 
Several alternatives for a closed form solution to the resection 
problem are given in literature. We follow the approach 
suggested by (Fischler & Bolles 1981). Named the “Perspective 
4 Point Problem” their algorithm solves for the three unknown 
coordinates of the projection center when the coordinates of 
four control points lying in a common plane are given. Because 
the control points are all located on a common plane the 
mapping in-between image- and object points is a simple plane- 
to-plane transformation T. The location of the projection center 
can be extracted from this transformation T when the principal 
distance of the camera is known. For a detailed description of 
the formulas please refer to the original publication. 
To complete the solution of the resection problem we also need 
the orientation of the camera in addition to its location. (Kraus 
1996) gives the solution for determining the orientation angles 
when the coordinates of the projection center are already 
known. The algorithm makes use of only three of the four 
points. 
In principle, the complete process, extraction of building 
silhouette, improvement of image coordinates by GHT and 
spatial resection can be repeated iteratively in order to avoid 
errors resulting from the initial orientation data. Nevertheless, 
for our application the differences between the projected wire- 
frame and the image were mainly caused by errors within the 
available model due to measurement errors or generalization 
effects. Subsequent iterations did not enhance the result 
significantly. 
4. EXPERIMENTS AND RESULTS 
A series of real-world examples was chosen from our database, 
to investigate both the feasibility of our approach and the 
accuracy obtained from photogrammetric processing. Our 
investigations are based on a dataset of the city of Stuttgart 
provided by the City Surveying Office of Stuttgart. This data 
was collected by manual photogrammetric stereo measurement 
from images at 1:10000 scale (Wolf 1999). For data collection 
the public Automated Real Estate Map (ALK) was additionally 
used. Thus, a horizontal accuracy in the centimeter level as well 
as a large amount of detail is available. 
  
Figure 4: Prototype of the mobile photogrammetry device con- 
sisting of a camera, a compass and a tilt sensor. The GPS is 
mounted separately. The laser unit was not used in this project. 
495