Topologically, these candidates are also compared with 
designed patterns to identify possible pieces of the feature lines. 
Therefore, this study designed 12 patterns (shown in Fig. 4) to 
compare with detected candidates and to extract possible pieces 
of the feature lines. In Fig. 4, the gray grids present possible 
locations of a section of one line. If certain grids do not match 
these patterns, the proposed process then directly considers 
them isolated noises and removes them. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 4. Designed patterns 
2.3 Registration 
After line detection, feature lines are extracted from the aerial 
image and LIDAR data for data registration. To estimate the 
correspondence between these two data sets, structure lines 
from the LIDAR data are back-projected onto the aerial image. 
The following step then transforms these two groups of feature 
lines into the Hough space. Because each building has a unique 
geometric orientation in the local area, this investigation 
considers the complete building structural lines to estimate data 
displacement. An illustrated example is displayed in Fig. 5, 
which shows the differences of parametric distributions 
between two buildings. This step is an iterative process for 
modifying the positions of projected structure lines and 
stopping the iteration when the displacement is smaller than the 
threshold. 
  
(b) 
   
(d) 
Figure 5. The difference of parametric patterns 
(a) Building I; (b) Building II; 
(c) Hough pattern I; (d) Hough pattern II 
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 
   
  
  
3. EXPERIMENTAL RESULTS 
This study selected the test area in Taipei City, Taiwan. In 2008, 
the LIDAR data were scanned with 10 points per square meter 
by using the Leica ALS 50. The original point spacing reached 
30 cm; therefore, the spatial resolution of PDSM is designed to 
reach 40 cm. The aerial image was captured in 2008 by a DMC 
camera. The spatial resolution of past images is approximately 
17 em. Figure 6 shows the aerial image and LIDAR points of 
two targeted buildings. Figure 7 shows the building boundaries 
before and after registration in the image space. For validation, 
the manually plotted corners were used to estimate the 
registration quality by root mean squared errors (RMSEs). The 
RMSEs in the directions of the sample and line reach 3-4 pixels. 
(b) 
   
(d) 
Figure 6. Past datasets of the targeted building 
(a) Aerial image in Case I, (b) LIDAR data in Case I, 
(c) Aerial image in Case II, (d) LIDAR data in Case II 
(b) 
  
  
    
    
  
  
  
  
    
    
   
  
    
   
    
   
  
  
  
  
  
  
  
  
  
     
   
    
  
Th 
the 
LI 
res 
cal 
res 
re; 
ini 
TH 
Cc