Paolo Gamba 
  
parts of the building start to be (erroneously) discarded, due to the small acquisition errors or noise in range 
measurements; this is clear, in fig. 3 (b) and (c), looking at the central and leftmost built structures. 
3 EXPERIMENTS 
The preceding technique has been applied to a number of different sites corresponding to urban structures in downtown 
areas. In particular, we refer here to some examples exploiting data over the town of Parma, Northern Italy. The dataset 
has been acquired on the town of Parma in June 1998 with the Toposys sensor installed on a plane of an Italian 
company called CGR, Compagnia Generale Ripreseaeree. The flight height was around 800 meters; the Toposys sensor 
is able to acquire, flying at that height, approximately five points per square meter, so that the one-meter grid which is 
usually delivered to the customers, and that we used, can be calculated with a good reliability. Up to now the Toposys 
instrument isn’t able to measure the reflected signal intensity, so it gives pure geometric data and it can acquire first 
pulse or last pulse alternatively: our data has been acquired in the last pulse mode. Nevertheless, thanks to the well 
equipped and powerful plane, we could also acquire, during the laser flight, aerial photogrammetric images that will be 
exploited in a further development step of our algorithm. 
The first example is depicted in fig. 4, and represents Piazza Garibaldi in Parma. We observe that the 3D shapes of the 
buildings in the area are much more delineated after the building extraction procedures than in the original data: all the 
significant features are considered, but the vegetated area in front of the tower building has disappeared, thanks to a 
sufficiently high similarity threshold (see previous section). Even the 2D shape (or better, footprint) of the buildings has 
been improved by the applied algorithm. 
     
50 
Figure 4. Data over Piazza Garibaldi, Parma, before and after object extraction. 
The second example portraits a different area in Parma, a residential area with a group of large houses and some trees. 
Even in this case the regularization of the 3D and 2D shapes of the artificial structures is evident. This is true especially 
looking at the small errors in the LIDAR DTM near the walls; after the building extraction procedure this kind of 
imprecision has almost completely been discarded. Moreover, the trees and bushes in the area have been either 
discarded or extremely regularized, so that their presence could be easily recognized with further simple processing 
steps. 
However, fig. 5 evidences also some problems in the procedure, already noted in Gamba and Houshmand, 2000, and 
related to the need to have a regular grid as the input of the analysis algorithm. When the searched objects are not 
exactly along the rows or columns of this grid, as in this case, we have possible reconstruction error, like those in the 
internal (with respect to the viewer) wall of the largest building. A second problem source is the use of a unique value 
for the threshold values in Table 1 through all the analyzed image. This is the cause, for instance, of the fact that the 
elevator small towers on the roff of the same large building are only partially considered in the output image (more 
precisely, the frontmost one has been discarded, while the one in the background has been considered). A possible 
solution to this drawback is the introduction of an adaptive threshold, taking into account the local regularity (to be 
somehow defined) of the area around each considered segment. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 317