Automatic construction of 3D maps of towns is a difficult problem since
urban areas can be very dense. Usual
techniques to achieve this goal rely on radiometric information, and use very
complex procedures involving a lot of parameters.
We have proposed in this report a new method to refine DEM.
This method actually gives an automatic way of extracting buildings
from a dense urban area by using a crude DEM as initial condition.
A first advantage comes from the kind of results given by the procedure
. Given a raster data, we obtain a vector which is useful
for semantic interpretation and compression goals.
Moreover this extraction is fast and efficient. It also provides robust
estimation of roofs.
The originality of this method can be appreciated as follows :
- First, the
object oriented approach we have chosen gives an elegant way of adding
geometric constraints during the extraction.
- Second, the way of adding a prior
knowledge as interactions between buildings gives a nice framework
to deal with poor quality data.
- Finally, since the data
term is built using a projection of a proposed silhouette on the data,
it is easy either to change the kind of data used in the energy term,
or to add a complementary energy term using another data.
However, this work can be improved by :
- testing the algorithm on more DEMs (optical, radar, laser etc...),
- improving the algorithm to be faster and to deal with larger areas,
- adding more complex models of buildings and roofs,
- improving the optimization step in order to be able to add a more
complex prior. Encountered problems are related to the high number
of local minimas of energy which appear when we add a better a priori knowledge
in order to deal with cruder data.
|