The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXX III. Part B5. Beijing 2008 
Then we can get: 
x = {B T PB) ' B T Pl (12) 
where: 
5 = 
0 
-Z-À.Z 
y + Ay' x + Ax ' 
-z-Az' -y-Ay' 
0 x + Ax' 
0 
l = {x'-x y'-y z'-zf 
So the solution of equation (11) X = (a,b,c) T can be obtained 
from the above equations. 
3. REDUNDANCY ELIMINATION 
After the accurate registration, large numbers of points are 
overlapped at the same place. So in each cube, choose the data 
of one station that the distance between the point and its own 
original scan-station is the shortest. At the same time, if the first 
part of data is not integrated, the partial data of another station 
which is secondly nearest to its own original scan-station are 
used to fulfil the hole in this cube. The other points in this cube 
are eliminated. Thus, the data which are remained at last are 
ensured to be valid and non-redundancy by means of this 
method. 
Data from two stations are in different coordinate systems 
before primary registration (see figure 5). And they are 
transformed into a uniform coordinate system basically after 
primary registration. However, two parts of data are layered, 
and the average error is 0.008m after primary registration. This 
kind of data can not be merged together and used directly. This 
is shown in figure 6. 
4. EXPERIMENTAL RESULT 
According to the related researches, a 3D application system is 
developed in VC++6.0 and OpenGL 3D graphic library to 
validate the feasibility and the accuracy of the algorithms. The 
point clouds obtained by Leica HDS4500, Leica HDS3000 and 
Leica HDS6000 are used in the experiments. Take two stations 
of data dngl and dng2 obtained from Leica HDS4500 for 
example. 
Figure 5 Data before Primary Registration 
Figure6 Data after Primary Registration 
(a) Data of girder layered 
(b) Data of pillar layered 
(c) Data of wall layered 
In the accurate registration, choose 1.5m as the side length of 
cube. After 7 times iterative calculations in 130.65s, two parts 
of data are registered very well. The average error is 0.00172m, 
and the standard deviation is 0.00194m. Layered data from 
different scan stations are registered with very little error, and 
they are not layered any more (see Figure 7). 
Before registration, each station contains 1220780 points and 
1222458 points respectively. After two steps of registration and 
spending 7.88s in redundancy elimination, there are only 
1578349 points when they are merged together. The data are 
reduced 35.4%. Thus, not only are the redundant data 
eliminated efficiently, but also the texture of point cloud are 
much clearer.