ra used in 
| 
ENT 
(umur) A«T'P1S 
x 
OM 
= 
in 
Ad 
  
; relative to 
noise level 
  
the use of 
type of 
—r—4 40 
35 
-| 30 
125t 
Bu 
+ 20 y 
das 
pny 
T 7] 10 
5 
Q 
  
  
is relative to 
jise level. 
) pixels. The 
good image 
r two as bad 
The camera 
se levels can 
| of imaging 
so tested in a 
creasing the 
much, unless 
ic network is 
re each new 
nation. In the 
1 adds only a 
ssential. This 
he projection 
ircular image 
he number of 
on upto 100 
rovement can 
  
40 T T T T T T T 40 
10 frames 
35 | 20 &ames ..... +35 
30 frames _ _ 
  
50 frames _._—. 
30 [7100 frames __... An 
150 frames m... 
25 | 200 frames ___..... 
N 
e 
T 
— 
Lan 
¥ 
> B 
(wm) APIS 
Std.Dev (mm) 
10 
  
  
  
  
Distance (m) 
Figure 3 Change in co-ordinate standard deviations relative to 
distance. Each curve depicts values with different number of 
images in image block. 
4. APPLICATIONS 
As mentioned in the beginning, this method is especially 
designed for measurements of large objects where imaging is 
done partly or entirely inside the object space. The concave 
parts of large objects are often troublesome in network design. 
This method provides help for such cases. What are then the 
application areas where we meet such conditions? In chemical 
industry there is need to update outdated structure models of 
complicated pipe structures, which may never have been built as 
designed. In such applications the visibility is the restricting 
element and this means that imaging must be done also inside 
the object space. Also all inner space modeling is a fertile 
environment for this method. The modeling can be 
measurements for the renovation or redecoration of an old 
building or reconstruction of remains of an ancient building in 
an archeological site. One application area can be found in 
forestry. In forest inventory, taking samples inside the forest is 
still a dominant method for estimating the annual growth of 
forest. Currently the volume of tree stems and scattering of trees 
in forest in samples areas are measured manually. This means 
measuring the diameter of tree stem and its distance from a 
fixed point. So, also in this case measurements are 
accomplished inside the object space and therefore this task is 
appropriate for the presented method. 
5. EXPERIENCES OF A REAL TEST 
Real experiment was accomplished in forest environment in 
order to test the capability of circular image block 
measurements in forest inventory application. For the test few 
targets were attached on the tree trunks. Otherwise natural 
object points were selected as for the tie points. In general only 
one tie point is needed per image to resolve the only unknown 
parameter of one image. In practice more observations are 
needed. It is to be noticed that all image observations are used 
for resolving the common unknowns of the block, namely 
rotations of the first camera pose and the length of the rod. That 
is why the distribution of image observations should be 
geometrically good on all images. 
From simulated examples you may have noticed that image 
measurements are supposed to be accomplished in sub pixel 
accuracy. In practice this was achieved by using template 
matching on subsequent images. On chosen source image 
template was extracted in size selected by operator. The best 
position of template on next image on sequence was resolved 
with largest cross-correlation. The final position was then 
estimated with LSQ-estimation with accuracy of sub pixel. This 
resolved best position of template on target image was then used 
to extract a new template on this image for matching on next 
image. So the template image and target image are always from 
subsequent images. This way we can be sure that viewing angle 
on these camera posies don't differ much. This semiautomatic 
matching continues until the point is out of sight or occluded. 
For matching stopping criteria some limit correlation value is 
assigned. 
This approach works well in one image block. Cause we need 
convergent observations we have to have same object points 
observed also from a second image block. In our 
implementation the operator can point out the measured point 
on the first block image and then point out an initial position on 
second block image for template matching. The rest 
homogeneous points are measured in a similar manner as with 
first image block. 
For image sequence measuring tool there was no adequate 
software found. So for all measuring tasks as well as for the 
estimation part, I had to create software of my own based on a 
Linux operating system. 
The estimation of the image block was not as straightforward 
computation as with simulated counterpart. The problem 
occurred with initial values. With non-linear estimation case 
good initial values for parameters are essential. With simulated 
data this was no problem. What was noticed with real data 
estimation was that if rotation angles o and « were near zero the 
¢ and a angles were strongly correlating with each other. Also 
parameter r and scale measurements had a great correlation. If 
initial values are not good enough the iterations process might 
convergence into a wrong local minimum solution. By using 
strongly weighted observations of ¢ and r-values the iteration 
can be directed into global minimum solution. As with 
simulated tests, observations only included normal distributed 
error, the camera model was supposed to be out of any 
systematic error components. With real data the influence of 
radial distortion not corrected seemed to be quite substantial. 
6. CONCLUSION REMARKS 
In this implementation template matching is used for tie point 
measurements. If numerous of circular image blocks have to be 
measured perhaps feature based matching procedure might be 
more practical than this semiautomatic measuring approach. 
Although in this first experience the most essential thing was to 
achieve an observation set without any gross errors than to 
create a more automatic procedure for image observation. From 
simulated tests and experiment with real data we can surely see 
that this novel method is applicable in those measuring tasks as 
it was designed for. 
REFERENCE 
Heikkinen, Jussi. Video Based 3D Modeling. International 
Archives of Photogrammetry and Remote Sensing. Vol. 
XXXII, Part 5, ISPRS Commission V Symposium 
Proceedings, June 2-5,1998, Hakodate, Japan, pp. 712-716. 
Heikkinen, Jussi. Circular Image Block Measurements. 
International Archives of Photogrammetry and Remote 
Sensing. Vol. XXXIII, Part 5A, ISPRS Commission V 
XVIII Congress Proceedings, July 16-23,2000, Amsterdam, 
the Netherlands, pp. 358-365. 
Heikkinen, Jussi. Video Measurements for Forest Inventory. 
SPIE Videometrics 22-23.01,2001. Vol. 4309. San Jose, 
California, USA, Spie -The International Society for Optical 
Engineering, pp. 93-100. 
24}