3. Istanbul 2004 
tting to reunite 
ze structure has 
ost. The octree 
sily (Samet, H., 
| merge process 
similar. Figure 
. However, one 
similar normal 
itional check of 
ring planes are 
ggered again to 
nd to ensure the 
e original lidar 
D planes. Each 
rameters (A, B, 
gular Irregular 
sualization use. 
in the form of 
parated ground 
ine successfully 
  
A and B (a) 
merge 
  
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
  
  
  
  
(b) 
Figure 7 Result of merged plane (a) border (b) TIN 
3. CLASSIFICATION 
A classification can be performed based on the attributes of the 
extracted 3D planes. In the time of preparing this paper, we 
have not developed a well classification method yet. However, 
simple classification results based on plane attributes such as 
area, average height, gradient, average intensity, shape, 
orientation and symmetry shows the potential of the feature- 
based classification. Figure 8a, b, shows the results of plane 
classification by area and gradient respectively. 
  
   
  
  
  
  
  
  
  
   
  
   
  
  
  
  
  
Figure 8. Results of plane classification by (a) area and (b) 
gradient 
Some classification examples of both airborne lidar and ground- 
based laser scanning data are shown in next section. 
4. EXAMPLES AND ANALYSIS 
The proposed method can be applied to both airborne and 
ground lidar data. However, the thresholds used in the program 
should be adjusted to fit different data sets according to the 
different scanning accuracy. Our test data include an airborne 
lidar data set collected in Hsinchu, Taiwan with Leica ALS40 
and a ground lidar data set obtained in The Eastern Gate of 
Hsinchu City with Optech ILRIS-3D laser scanner. Table 1 lists 
the basic data attributes and the thresholds applied in the tests. 
  
ALS40 ILRIS-3D 
  
Scan Date April 14 ,2002 February 16, 2004 
  
The Eastern Gate of 
Scan Area Hsinchu : € 
Hsinchu City 
  
Scan Accuracy 30 cm 0.7 cm 
  
  
  
  
  
  
Point density About 2.3pt_ About 907pt/ 
Point cloud size 515,991 464,563 
Distance 
threshold 1m 0.05m 
Area threshold |1/4 node face area| 1/4 node face area 
Angle threshold 30 3 
  
  
  
  
Table 1. Information of examples 
4.1 Airborne lidar example 
Figure 9.a shows a set of point cloud of airborne lidar covering 
an area of 500 x 500 m?. Most of the points distribute densely 
over the ground surface and top of buildings, and points scatter 
on side wall of buildings are fewer. Figure 9.b shows the border 
of best-fit planes after the split process. Table 2 lists the 
computation time, the parameters of the octree structure, and 
the statistic data of extracted planes. In this case, 96.7% of the 
total points were used to form planes, the rest are scattered 
points which cannot be used to form planes. Figure 9.c shows 
the border of merged process. The stadium ground to the west is 
merged from pieces of planes. Figure 9.d, e, f, shows the results 
of merged plane classified by area, gradient, average intensity 
respectively. In Figure 9.e, vertical planes (building wall) and 
horizontal planes (building roof and ground) can be classified 
by their gradient.