International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
detection’ approaches are proposed in this paper. À first 
approach is to find candidates of building shape changes, 
especially height changes, for large buildings. À second 
approach is to detect candidates of building's shape and 
texture changes for small buildings. 
5.1 Coarse change detection 
Candidates of building shape changes are detected by 
comparing DSM (or height information) of existing 3D 
building data with DSM generated from TLS image. 
Coarse changes in urban areas are detected, which is 
probability of shape change. This preliminary 
information is assigned to each building. 
In fact, when there is a building demolition, a building 
height is changed from an existing level to a ground level. 
When a new building is built up, the building height is 
changed along an opposite direction. Moreover, when a 
building is reconstructed, there are some changes of 
building height. 
This method is effective to detect changes of large 
buildings such as department stores or multi stories 
parking lots, etc. in a viewpoint of processing time. 
However, most of small building changes cannot be 
detected since they are built on full lot size due to very 
small lot area. Therefore, fine change detection is applied 
to all buildings including these small buildings. 
5.2 Fine change detection 
Existing 3D data are projected into TLS images. Building 
changes are detected by using changes of textures in the 
TLS images and the preliminary information. Not only 
roofs but also walls are referred in this processing. 
Building shadows are extracted from TLS images by 
using the temporal information of the data acquisition. 
TLS images are enhanced not to influence the texture 
change detection. Details are described in the followings. 
I) Decomposition 
Existing 3D data are decomposed to roof polygons and 
wall polygons as Figure.7 shows. A classification of 
roofs and walls is done based on the roof model, which 
represents buildings by using roof boundaries, in the 
SNAKE-aided 3D data refinement processing. 
2) Back-projection 
These polygons are back-projected into visible TIS 
images; since occlusion, which is caused by surrounding 
buildings, can be guessed by a geometric processing. 
Generally, roof textures are back-projected to a TLS 
nadir image, and wall textures are back-projected to TLS 
forward / backward images. When non-textured polygons 
exist or TLS images do not have corresponding areas 
with polygons, they are not back-projected to TLS 
images. 
  
   
    
  
  
Existing 3D data 
     
ee € 
"WA m 
(7 
‘Decomposition’ NT 
  
  
  
  
Wall polygons 
Roof polygons 
  
  
  
  
  
   
Forward Nadir image Backward 
image image 
  
  
  
Figure. 7. Texture reference 
3) Polygon based area correlation 
A polygon based area correlation is calculated between 
existing 3D data and TLS images. An equation of the 
polygon based area correlation including parameters is 
described in eq. 1. 
Ctotal = F@ Cr -fà* Cw) 1 ER FQ) (eq. 1) 
Nr 
Cr =3 (Crk) Nr 
kzl 
Nw 
Cw =) (Cwk)/ Nw 
k=l 
where 
Crk : Correlation coefficient in each textured roof polygon 
Cwk : Correlation coefficient in each textured wall polygon 
Nr : Number of textured roof polygons 
Nw : Number of textured wall polygons 
fÔ,/@; Weight parameters 
Ctotal; Result value for for change detection 
A theory threshold value of a correlation coefficient in 
building change detection is not approximately 1.00. 
1326 
Interna 
Howe 
cables 
nadir 
chang: 
prelim 
detect! 
demol 
correl: 
when 
Figure 
up in 
detecti 
In this 
be use 
  
A 
| 
| 
yi 
| 
| 
| 
| 
| 
| 
| 
I 
I 
| 
| 
I