in 
ce 
of 
ta 
ht 
ic 
in 
nt 
an 
ur 
1g 
Or 
231 
3.2 Operator controlled measurement 
The image matching results have been controlled interactively at a digital photogrammetric stereoworkstation 
(Intergraph Imagestation 6787). A certain problem for interactive work at digital workstations is the relatively 
low precision of manual measurements of about 1/3 of the pixel size. A way out of that problem is to execute 
image matching with digital images of a coarse resolution and to use a fine resolution of the image pair for the 
control measurements. In the experiments the 120 um pixel size was chosen for matching and the 30 pm pixel 
size was used for the interactive control. Because our main interest is in a quality comparison of different colour 
matching solutions this seems to be an acceptable procedure even though a dependency on the pixel size can 
not be reasonably investigated in this way. 
In the investigation about 120 points were selected in each project. The selection of the points is done by the 
operator. He is choosing points which in his opinion are well suited for the interactive work, i.e., which are 
easy to identify and measure in the second image. Suitability for matching was not taken into account in this 
step. Some intensitity based correlation technique (included in the Intergraph software) is used to improve the 
manually measured corresponding points. The success is directly verified in stereo and, if necessary, manually 
corrected. By relative orientation a precision a, of 4.2 um was estimated for the operator controlled measure- 
ment in the Glandorf project. This is about 1/7 of the 30 um pixel size and 1/30 of the 120 um pixel size. 
This precision is sufficient for having appropriate check values for the colour image matching experiments (at 
the 120 um pixel resolution level). 
3.3 Theoretical precision 
In the investigation our main interest is in the comparison of different possibilities of matching colour images. 
Theoretical precision of matching is calculated by the covariances of the estimated parameters. In the discussion 
in section 2 we outlined that the best theoretical precision is expected for the simultaneous multichannel solution. 
The five alternatives taken into account are matching each colour channel R,G,B separately and matching the 
intensity components I, Y calculated by the colour transformations mentioned above. In the analysis we restrict 
ourselves to the precision of the estimated parallax of the selected points. If a local coordinate system (z, y) is 
used which is centered at the point to be matched the estimated shift parameters ao, a3 are just the parallaxes 
pz,py in that point. 
Many experiments in area based image matching have shown that the theoretical precision of estimated paral- 
laxes is often too optimistic by one order of magnitude. The larger the matching area the more unrealistic is 
the estimate. Reasons mainly stem from neglecting correlation effects emerging, for example, from data capture 
and resampling. Therefore, we focus on the relative precision between the different matching variations. The 
quotient of v, ,[R] of the red channel to c, [multichannel] is listed in table 2. The values obtained for the other 
components [G;B;I;Y] are also included in this table. op; is determined by o4 /qa, with the a priori noise c,, and 
the cofactor qa, which is the corresponding diagonal element taken from the inverse normal equation matrix. 
'The same applies for the y direction. The ratios of estimates of noise in the different solutions are also shown 
in table 2. The estimated noise increases approximately linear with growing window size. The noise estimates 
for the multichannel solution range from 2.4 to 4.0 in the Glandorf project (window sizes 11 x 11 to 55 x 55 
with increments of 2 pixels) and from 3.5 to 5.6 in the IR-imagery of the Schorndorf project (window sizes 11 
x 11 to 63 x 63 with increments of 4 pixels). The relative precision estimates (ratios of parallax precision) are 
nearly constant for all window sizes. Thus, the mean values of the estimates collected over all points and all 
window sizes are listed in table 2. 
  
  
  
R:mc G-:mc B:mc I:mc Y:mc project sample 
px: 1.41 1.76 2.54 1.87 1.74  Glandorf 122 pts. 
py: 1.41 1.77 2.55 1.88 1.75 
noise: 1.12 1.00 0.85 0.84 0.91 
px 1.67 1.79 4.45 2.30 2.03 Schorndorf 120 pts. 
py: 1.64 1.84 4.58 2.30 2.07 
noise: 1.19 1.04 0.56 0.83 0.92 
  
  
  
Table 2: Theoretical precision and noise: a comparison between the different solutions of matching colour images 
(me: multichannel). 
The decrease in the theoretical parallax precision from red, green to blue shows up in the numbers of table 2. 
Significantly best for matching is the red component of the RGB colour image and the IR component of the 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995