Michel Morgan 
  
3.3.2 Face extraction/ vectorization 
  
50 m 
; du: ; : segm 
A least squares adjustment for plane fitting is used based on the height values of each of the main roof faces before 2). t 
applying the majority filter (to make sure that there are no gross errors or blunders in the height values). After applying n Sf 
the iterative majority filter, the adjacency among the roof faces can be obtained. The way of vectorizing the fac = 
depends on the adjacency among faces. Three 
acia ai M iol ; : secon 
If the face is adjacent to only one or no other faces in addition to the background, primitive matching (matching the face buildi 
segment with one of the geometric primitives and extract the primitive parameters) is done for that face. The geometric the fi 
primitives can be chosen according to an inventory of roofs in the area of interest. In this case the planimetrc three 
information is used to obtain the primitive parameters and the height information is added to the primitive based on the buildi 
plane equation of the face. Primitive matching can be done based on the analysis of the compactness and the relative the fa 
moments of area along the principal axes of the face. However, because of the time limitations, primitive matching wag based 
not used in the present research. Instead, only rectangle extraction is used based on defining the principal axes and adjace 
matching the moments of area along the principal axes with the parametric rectangle, followed by extracting the Consi 
rectangle parameters. Model matching had been done before in Maas, 1999), (Weidner and Forstner, 1995) and lines ( 
(Weidner, 1996). However, in this research matching is done for each roof face to consider more complicated roof averas 
t 
structures. groun 
If the face is adjacent to more than one face, the outline of the face are extracted based on the intersection of the fac 
and the adjacent face. For the faces adjacent to the background line fitting is done to obtain the respective edges. 
4 EXPERIMENTS AND RESULTS 
The laser data is re-sampled into regular raster with a pixel size of 0.577 m which is the invert of the square root of the 
maximum density of the laser data (3 points per square meter). The data are re-sampled twice; using nearest neighbor 
interpolation, which is used for building detection and using bilinear interpolation, which is used for differentiating 
between buildings and vegetation and for building extraction. The procedure for building detection and extraction is 
done for two test areas. The DSMs of the test areas after re-sampling into regular raster are shown in figure (1). The 
DSM is represented as an image, elevation values converted to gray values. 
E 
m 
E: 
eM 
m: 
um 
- 
m 
(a) (b) « 
Figure 1. The DSM after re-sampling the laser data into regular raster for (a) the first and (b) the second test area 
Applying the median filter before the connected component labeling re-classifies some of the misclassified pixels o 
noise inside/outside the objects. Moreover some of the thin lines connecting segments are eliminated. The eliminatio! Fic 
of those thin lines is important in the case where 8-directions connectivity is used for segmentation. The number 0 the 
segments is smaller than that without applying the median filter. The minimum building area used for classification i 
  
620 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.