derives disparity updates and uses a varying window size 
in order to minimize the uncertainty of the disparity. 
6.3 Interior Orientation and Model Formation 
For the interior orientation, the fiducial marks on 
the image frame are digitized by the operator. 
Coordinates are transformed from the pixel coordinate 
systems to the image coordinate systems by affine 
transformations. The formulae for this affine 
transformation are given, e.g. by Moffit and Mikhail 
(1980, pp. 291-295, pp. 592-593). 
As the images have been taken with a terrestrial 
photogrammetric camera, the parameters of the relative 
orientation are known. The image coordinates of 
corresponding points which determine the disparities d, 
and d, as 
dix -xn (6.1) 
dy = yL - JR 
can be used according to the collinearity condition to 
compute object space coordinates. 
The formulae for the computation of spatial 
coordinates X, Y, Z according to the collinearity 
condition are given e.g. by Kraus (1982, Chapters 2.3 
and 4.11). 
Final results of the matching method are the spatial 
coordinates of points on the surface of the object. They 
can be used for further applications. 
6.4 A Set of Uncertainty Measures 
In order to estimate the quality of the matching 
results, there is a need for a reliability measure. The 
problem to provide such a measure for a vision 
algorithm that encounts both for accuracy as well as 
robustness is known to be hard (Forstner and Ruwiedel, 
1992). In this method, a measure which is meaningfully 
including all aspects of the matching (e.g. edge detection, 
feature matching and subpixel-accurate point matching) 
in a coherent numeric system cannot be provided. Instead 
of that a set of uncertainty measures each making a 
statement about the outcome of a specific step in the 
processing is available. 
The match function (5.1) gives a similarity measure 
s about the overall similarity of the two edge segments 
compared. The higher the similarity measure the more 
accurately the edge segments correspond. In this sense a 
high s-value can indicate a high probability of correctly 
matched edge features. Nevertheless, it does not prove a 
correct match. Correctly matched edges will most 
probably have a high similarity value. However, if the 
object has several similar features, e.g. regular patterns, 
wrongly matched features will also obtain a high s-value. 
To draw conclusions about the correctness of a 
match becomes much easier when in addition to the 
similarity measure of feature pairs the strength of a 
connected component is considered. When many matches 
have been found during the propagation of a hypothesis, 
the probability that these matches are correct is high. 
This property is used in order to solve ambiguities 
occurring between different hypotheses. 
The measure available to assess the quality of a 
matched pair of points is the uncertainty of the disparity 
update O44 as a result of the area-based matching 
algorithm of Kanade and Okutomi (1990). With the help 
of oq the accuracy of the object space coordinates can be 
estimated. 
Summarizing, it can be stated that the similarity 
measure of the feature matching in combination with the 
strength of the connected component allows conclusions 
about the reliability of the matching, whereas the 
uncertainty of the disparity mainly contains information 
about the accuracy of the matching. A more 
comprehensive investigation about the robustness of 
object space coordinates derived from matched edge 
features was done e.g. by Faugeras et al (1992). The 
derivation of meaningful reliability measures directly 
resulting from a matching method without any 
comparison with results of other methods or true data has 
to be left to future research. Methods and approaches to 
solve parts of the problem can be found in Fórstner and 
Ruwiedel (1992). 
7. Implementation and Experimental 
Results 
The developed feature-based matching method was 
implemented on Sun workstations using the X-Window 
user interface and applied to imagery for car accident 
investigation. Final goal of the measurements was the 
determination of the depths of dents in the car bodies. 
7.1 The AMORPH Stereo Vision Software 
Package 
The amorph stereo vision system provides a digital 
photogrammetric software package. Although its main 
components are determined by the implementation of the 
proposed stereo vision method, it contains the tools to 
interactively perform digital photogrammetric 
operations, such as visualization and modification of 
digital images; input, selection and subpixel matching of 
corresponding points; interactive verification and 
correction of the matching results; interior orientation 
and model formation as well as selective output for 
further processing. Thus, the software can be used 
without depending on the developed stereo method. The 
main areas of application of the AMORPH software are 
basic digital photogrammetry and the use as an 
environment for stereo vision research. 
The AMORPH software is designed in a strictly 
modular manner. General enhancements as well as 
modifications of the graph-based stereo vision method 
(e.g. based on recent research results) are easily 
implementable. 
7.2 Accident Investigation Imagery 
Images of cars are characterized by well-determined 
curved edges of various shapes as well as specular 
reflections in some parts. All images were taken outdoors 
under common conditions for close-range 
photogrammetry. No special measures were observed. 
The images were taken with a 24 x 36 mm? camera on 
colour slide film. They were digitized with an Apple 
McIntosh slide scanner using the DeskScan software for 
the FrameGrabber card. The images have a resolution of 
about 600 x 400 pixels. 
The first picture shows the image pair of a grey car 
(Fig. 7.1). The damaged door of the car is used as the 
object of demonstration. 
The generated edge neighbourhood graph can be 
314 
see 
unc 
gro 
reg 
line 
are 
par 
mat 
edg 
unn 
Fig 
inte 
pro 
Uni 
unc 
SOIT 
Ho 
exti 
prox 
abo 
foui 
devi 
for 
iden 
vari: