228 
estimation of a certain object in an image, for instance a traffic sign, is another purpose of matching. Template 
matching, of course, is not a very flexible solution for identification and location of general 3D-objects in images. 
Establishing a reasonable library often restricts to a narrow application field, for example, in machine vision 
inspection. Taking different clues into account for identification and location of traffic signs is the main topic 
of Geiselmann and Hahn (1994). In this study the contribution of the sources colour, motion (in an image 
sequence) and invariants of object shape for identification has been investigated. 
Nowadays in photogrammetry a high percentage of the airborne photographs are aerial colour images. In 
principle, photogrammetric tasks like point transfer or data capture for DTM reconstruction and others can be 
carried out with colour images. For this tasks matching algorithms may be used which assist a human operator 
in his daily practice. Looking closer to practice shows that all this basic tools restrict to grey tone or one channel 
images. Therefore a pragmatical solution would be to use just one of the RGB channels. To our knowledge no 
investigation in matching aerial colour images is reported. 
The estimation of optical flow in image sequences can be considered as an image matching task with fundamental 
similarities to image matching in stereo vision. Markandey and Flinchbaugh (1990) have taken multispectral 
data into account for optical flow computation. Experimentally they found that the multispectral solution (with 
colour TV data) was somewhat more accurate (around 20%) than a single spectral solution. The numerical 
stability of the flow field equation improved considerably from what they concluded that multispectral optical 
flow is fairly robust. 
The scope of this investigation is to present a formulation of a simultaneous multichannel solution of area based 
image matching. The classical model is generalized using a vector valued image function. Experimentally we 
study quality aspects of colour matching and compare it to the results of alternative solutions. One alternative 
is to restrict matching, for example, to the red channel. Another procedure could be to match the red, green 
and blue channels of a stereo pair separately and fuse the results. As a third possibility we consider the RGB 
transformation to other colour spaces taking the intensity (brightness) data for matching. 
Questions which are behind this study are mainly of practical nature. Today there is a general trend from black 
and white to colour photogrammetry. An interesting point is that even economic aspects support this trend. 
For example, Colomina and Colomer (1995, p.37) report on map production using colour orthophotos. They 
observed a faster and better identification which increases the overall productivity by 6 % if colour orthophotos 
are used instead of grey tone orthophotos. Questions we keep in mind with respect to image matching are: 
Which quality differences can be expected theoretically between a multichannel solution and a solution with 
one of the channels? What show the experiments? What are the reasons and most important factors? 
2. AREA BASED MULTICHANNEL IMAGE MATCHING 
The formulation of area based image matching can be found, for example, in the textbook of Haralick and 
Shapiro (1993). Restricting matching to small image patches of two images the nonlinear model reads as 
Ti(z, y) — h[Is(p, a) * n(z, v) (1) 
with the geometric and radiometric transformation between the windows 
p(r,y) = ag-a£-ray 
qg(r,y) = aà3+a,x+ay 
h(I5) u agl» Ta. 
The geometric model for mapping from image to image is described by an affine transformation. For compensa- 
tion of radiometric differences between the two windows contrast ag and brightness difference a7 are introduced. 
The observational noise n relates to the intensity differences between the images I, and Is. 
Given approximate values for the parameters a, i = 0, 7 by linearization of (1) the following model is obtained 
dag daa T 
A! = ae V I»(p, q; aj) ( 1 2 y ) da; das + Ta(p,4; a;)das + 1da7 + n(x, y) (2) 
das das 
with 
AI = H(x,y)— h(I»(p,4;a?)] 
8. +0 
= —I5,—1l5). 
V I5 Gp 23 dq 2) 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995