3b. Beijing 2008 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
407 
building in this 
eral, presumes 
: kind of object 
as abstractions 
nportant to find 
i.e. the results 
n terms of the 
ition (Streilein, 
istraints can be 
model. A rigid 
lich has to be 
i these models 
which can be 
with the large 
utilized model 
buildings are 
it lines, in our 
al polyhedron, 
aption that the 
ind the borders 
nd plan. This 
instruction of 
eal with very 
describe the 
tion (BRep) is 
intation. Many 
l properties or 
le object is 
1 into a set of 
; additionally 
tow the faces, 
n constructive 
combined by 
itation always 
) a BRep no 
guarantee the 
>n 
one or more 
> of a cuboid 
lip roof. This 
r (Englert and 
lex buildings, 
ed into these 
tically by the 
the result of a 
ctures. Every 
ice position, 
orientation and horizontal extension of each cuboid is already 
defined by each rectangle, only the height of every cuboid as 
well as roof type and roof slope have to be determined as 
remaining parameters for the building primitives. 
The parameters of the building primitives are estimated by a 
least squares adjustment, which minimizes the distances 
between the DSM surface and the corresponding building 
primitive, i.e. the building primitives are fit to the DSM surface. 
In order to apply the least squares adjustment first the 
appropriate model has to be selected. Additionally roof regions 
which do not fit to the selected model have to be excluded from 
the least squares adjustment to avoid gross errors of the 
estimated parameters. For both tasks the result of a 
segmentation of the DSM are used. This DSM segmentation 
into planar surfaces is supported by introducing ground plan 
information. Of course the given ground plan restricts the 
extension of the DSM area which has to be examined.More 
important, the implemented segmentation within each ground 
plan area can be based on the direction of the surface normals 
of the DSM, since possible orientations of planar surfaces to be 
extracted are predefined by the outline of the building. This is 
motivated by the observation that the direction of the unit 
normal vector of a possible roof plane emerging from an 
element of the ground plan has to be perpendicular to this 
segment. Hence, the different segments of the ground plan 
polygon are used to trigger the segmentation of a planar surface 
with a projected normal vector perpendicular to this element. A 
more detailed description of the segmentation process can be 
found in (Haala et al., 1997). 
4. SYSTEM INTEGRATION 
We develop a system to integrate all the above method, The 
reconstruction of buildings is split into four steps: 
in the first step the origin stereopair are preprocessed to 
generate the DSM data. Then a detection of the image edge is 
carried out. In the third step, a coarse model is recovered, which 
consists of the three-dimensional edges of the building. In the 
final step, the coarse model is projected to the original images 
to refine the model. In the following the four steps are described 
in detail. 
4.1 Stereopair DSM generation 
Because of the huge dense match cost of graph cut algorithm, 
we provide a batch process on the DSM generation of all the 
images, constructing the Disparity Space Image (DSI), then the 
energy minimization framework, minimized by the graph-cut 
algorithm, is taken to generate the corresponding DSM data. 
The generated DSM data contains the relative initial three- 
dimensional coordinates of the building. 
4.2 Edge detection of building rooftop 
The second step, by selecting the appropriate parameters, the 
image segmentation is applied with the EDISON algorithm, 
which separates the building areas from the background areas. 
Then the EDISON algorithm is used again to detect the building 
edges of the stereopair. After removing the noise of edge 
segment, line fitting and end points competition are introduced 
to optimize the edge. The BEFV is generated after the 
optimization. 
4.3 Coarse model extraction 
In this step, the coarse building models are generated according 
to the DSM data acquired from the first step. The left image of 
the original stereopair is chosen as a part of the outer interface 
to the user, so that the user can clearly recognize the building 
area. In the course of the practical implementation, on the outer 
interface to the user, the only operation that the user needed to 
do is to click the mouse in the interior of the building areas. On 
the inner interface, the corresponding areas in DSM data is 
selected, then based on the selected areas, the BMD algorithm 
is applied to produce the coarse building model. The property 
of the BMD algorithm is described as follows: Automatic 
building roof planar separation algorithm. Normal of each point 
in DSM is used as voting to detect planar. Flowing by isolate 
point assignment, planar joining can optimize the detected 
planar. And multi-thresholds are used to make algorithm more 
robust. Planar edge points competition strategy is an important 
mechanism which can eliminate the bad influence caused by the 
order of planar detected out. The ridge lines are finally detected 
by intersection of two planar. The coarse building models 
possess the initial attributes of the building, such as the length, 
width, height, footprint, the ridge line of the hip-roof building. 
4.4 Refine coarse model 
The final step, using the BEFV generated in the second step, the 
coarse model could be refined, project the edge of the coarse 
three-dimensional model to the original stereopair image to 
produce a line buffer in the corresponding BEFV, the WLDC 
algorithm is proposed to find the corresponding line segment in 
the stereopair, the WLDC algorithm is described as follows: To 
get the direction of the optimal line segment, the length and the 
direction of each line segment in the line buffer is taken as the 
parameters of a weight function: W=f(l, d), the line segment 
with bigger W has bigger influence on the optimal line 
segment's direction. A pair of corresponding optimal line 
segment is generated after executing the WLDC algorithm. 
Then the precise three-dimensional coordinates of the building 
edges can be derived by a space intersection, the coarse model 
of the building can be refined to a more precise three- 
dimensional model. 
5. TEST 
We have tested our integrated system .A complete process 
introduced in section four has been made on a stereopair. 
The stereopair in Figure 1 is taken as the input data to test the 
integrated system. 
For the stereopair in Figure 2, we preprocess it under the energy 
minimization framework using graph cut algorithm, a huge but 
acceptable computation will be carried out. Then we display it 
in a disparity map image type in Figure 3. The result of the 
edge detection using Edison algorithm is in Figure 4. After 
these, project the dsm data into the edge map to refine the 
coarse model generated in 4.3.