ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
orientation data introduce some uncertainty in the location of 
the epipolar line, stretching it to an epipolar band, whose 
width depends on the accuracy of EO parameters. The overall 
image sequence processing workflow, including the real-time 
components as well as the post-processing part, is shown in 
Fig. 3. For the single image approach, the buffering of the 
previous image and the complete stereo processing module 
can be omitted. 
The above design was initially prototyped in Matlab and 
since then has been implemented in Visual C++ environment. 
The low-level image processing steps that are identical for 
both stereo and single image processing are discussed in 
detail in the Cairo paper; here there is only the end result of 
the processing shown in Figure 4, illustrating the robustness 
of the method on a rather difficult road surface. 
  
  
  
  
  
  
  
  
  
| Data Acquisition Image n-1 Image n Relative Motion 
= smaiss nmm 
RGB to S RGB to S 
  
Transformation 
| 
Transformation 
| 
  
Image 
Preprocessing 
  
  
  
Median Filter Median Filter 
Binary Conversion Binary Conversion 
| ] 
Boundary Points Boundary Points 
Centerline Extraction 
  
  
Centerline Extraction 
  
  
  
  
Stereo Image Processing 
  
    
  
  
  
Figure 3. Real-time image processing and post-processing workflow. 
To assess the feasibility of the real-time image sequence 
processing in stereo mode, several simulations as well as real 
tests were performed. For example, for real-time EO 
parameter accuracy of 3 cm and 0.5° respectively, the image 
matching based on a 20 feature point model, executed on a 
dual Pentium 4 at 1.7 GHz is about 0.26 s, allowing for a 
maximum image rate of 4 Hz, while errors in EO of 0.5 cm 
and 0.1°, respectively, allow for image rate of 21 Hz, since 
the total image matching time is only 0.048 s (Toth and 
Grejner-Brzezinska, 2001a). Obviously, the image rate can be 
increased by using a smaller number of feature points such as 
a 6-10 point model, which will result in a 10 Hz rate at the 
coarser EO accuracy however, the robustness of the process 
can be impacted. Based on simulation tests, the low-level 
image processing tasks up to the polyline extraction seem to 
execute fast enough for real-time execution. The more 
complex image matching, however, may be a less realistic 
target for full real-time implementation. 
A - 364 
4. POSITIONING ACCURACY OF THE 
NAVIGATION MODULE 
The details of the concept of the use of real-time 
navigation data to support on-the-fly image processing are 
presented in (Toth and Grejner-Brzezinska, 2001a and b; 
Grejner-Brzezinska and Toth 2002). In this paper, we only 
present the estimated real-time navigation accuracy to 
support the claim that the currently implemented hardware 
(Litton LN100) provides sufficient short-term accuracy to 
enable image matching in real time. In other words, the 
change in position and attitude between two image captures 
can be estimated at the accuracy level that the image 
matching can be achieved at real-time. In order to 
demonstrate this accuracy, the reference solution based on 
GPS/INS data was used as a ground truth, while free 
navigation solution provided the actual real-time trajectory. 
Since two consecutive images collected at (typically) 10 Hz 
rate (which allows for about 1.3 m overlap), are matched 
using RO parameters provided by the real-time navigation 
solution, the primary accuracy measure for RO is the epoch- 
to-epoch rate of change of position and attitude error 
estimates. In practical terms, if two subsequent images are 
similarly misoriented (i.e., contain a similar amount of error)