-B3, 2012 
inal point clouds 
rameters. 
rs 
the Netherlands. 
te the detection 
ling had several 
second building 
\R data used in 
2007. The point 
"ure 5 shows the 
both buildings. 
mately 20 cm, a 
SM during the 
sed in this study 
m. This value is 
apet height. 
  
ise II (Unit: m) 
shown in Figure 
els denote the 
transform was 
ges into several 
ch group of line 
ients. Figures 8 
es developed in 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
(a) M mU 
Figure 6. The detected PDSM structure lines for 
(a) Case I and (b) Case II 
  
(a) (b) 
Figure 7. The PDSM structure line segments for 
(a) Case I and (b) Case II 
  
(a) (b) 
Figure 8. The 3D structure lines for Case I in (a) the first view 
and (b) the second view 
  
(a) (b) 
Figure 9. The 3D structure lines for Case II in 
(a) the first view and (b) the second view 
This study also used octree-based segmentation (Wang and 
Tseng, 2010) to compare and evaluate the capabilities of TEA. 
The same point clouds were employed to derive coplanar points. 
  
  
  
The segmented surface points are shown in Figure 10 and are 
used in the planar equation to calculate the coefficients. The 
used thresholds contain the initial grid size, split distance, 
merge distance, and minimum point numbers of each patch. 
These values are manually optimized for the line computation. 
The computed structure lines were then obtained from the 
intersected planes. Finally,. the position differences were 
compared to calculate the root-mean-square errors (RMSEs) of 
the three axes. The comparison results are shown in Table 1. 
  
(a) 
    
M 
Figure 10. The octree-based segmentation results for 
(a) Case I and (b) Case II 
(b) 
  
  
  
Unit: m RMSE-X RMSE-Y RMSE-Z 
Case I 0.354 0.331 0.406 
Case II 0.351 0.340 0.468 
  
  
  
  
  
  
143 
Table 1. Comparison results of relative accuracy 
4. CONCLUSIONS 
This study proposed a method that employs TEA to identify 
three-dimensional structure lines. The two criteria of this 
method are the grid spacing and elevation difference, which 
were set as 0.25m and 0.5 m. Note that the grid spacing can be 
computed from the used LIDAR data. This value is lightly 
larger than the average point density. The independent structure 
lines were identified and compared with the results of the 
octree-based split-and-merge segmentation algorithm. The 
RMSESs of the two cases show that the differences between the 
three axes reached 0.35 m, 0.34 m, and 0.46 m. Case II had a 
greater elevation difference, which may be because of the 
parapet on the rooftop. The height of the parapet was 1.0 m.