bul 2004 
the case 
ange in 
: sudden 
onse. 
ased on 
letection 
letection 
  
in very- 
ed ridge 
y 
5. 
  
pothed 
riginal, 
e 
4.0 Junction detector using improved ridge detector 
We applied our technique on an extract of a IKONOS 
panchromatic scene above Ghent. The extract is taken over a 
suburban area containing 38 road junctions (cfr Figure 2). On 
this extract, ridge detection was performed using a window size 
w=9. Thresholds were placed on the gradient magnitude and 
eigenvalues to optimize the detection rate (i.e. true positives 
versus false positives). Region growing was then applied to 
improve the result in the vicinity of junctions. We found that 
performing region growing on a smoothed version of the image 
instead of the original image improved the later steps. 
Especially the thinning procedure which is necessary to produce 
a vectorized result can be sensitive to noise in the detected 
boundaries of the road. 
Figure 7 shows an example of the regularizing effect of 
(gaussian) smoothing on the vector result. Figure 7a shows the 
junction in the original image. Figure 7b the result of region 
growing and the thinned pixel chains using the intensity values 
of the original image. Figure 7c shows the result using the 
intensity values of a smoothed image. The latter shows 
smoother road boundaries which leads to lesser artifacts in the 
vector result. 
Region growing is applied with an adaptive threshold 7-7. 
Figure 5 shows a typical example of the detection result. To 
improve the robustness of the detection, we can filter out the 
junctions which belong to road segments of a certain minimal 
length. Figure 8 plots the detection rate in function of the 
minimum segment length. For a minimum length of 20 pixels, a 
true positive rate of 70% and a false positive rate of 12% is 
achieved. Increasing the threshold lowers the false positive rate 
below 10%. 
100 
80 
60 
  
% 
40 | 
gout Teall 
0 ; = 
10 15 20 25 30 35 
min segment length [pixels] 
  
Figure 8. Detection rate versus minimum segment length. Full 
curve gives true positives. Dashed curve gives false positives. 
5. DISCUSSION 
In this paper, we presented a junction detector based on an 
improved ridge detector using region growing. The detector is 
specifically tuned towards detection of road junctions in very 
high resolution images, where we observe a clear deviation of 
the simple line model especially in the vicinity of junctions. 
Region growing offers a simple model to extend the 
performance of the basic ridge detector. Of special interest is 
the possibility to include spectral information in the detection. 
Whereas ridge detection is a purely geometric detector related 
to image structures (as is the case for e.g. a gradient operator), 
region growing allows us to include spectral information within 
the road model. In this respect, a critical parameter which has 
819 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
not yet been discussed in this paper is the adaptive threshold 
used by region growing. Within our experiments, this threshold 
has been set manually. In further work, the link with the 
spectral properties of the road class will be investigated through 
the use of supervised classification. 
The experiments as presented in this paper show a reasonable 
detection of junctions. More importantly, the number of false 
alarms can be kept low. This is essential in our work towards 
the use of image derived information for automated quality 
assessment of GIS data. The graph matching technique that is 
applied to correspond the image data with the vector data is 
able to detect a certain amount of false alarms, but keeping this 
number low is essential for a reliable result. 
ACKNOWLEDGEMENTS 
The authors wish to thank the support of the Federal Science 
Policy. This work is performed under the project "The GIS 
problem detector" which is part of the STEREO research 
programme. Special acknowledgement goes to Tele Atlas for 
the use of their dataset. 
REFERENCES 
Arcelli C., Cordella L. P., Levialdi S. (1981) From local 
maxima to connected skeletons. IEEE Trans. Pattern Anal. 
Mach. Intell. 3(3):134-143. 
Beaudet P. (1978) Rotationally invariant image operators. In: 
Proc. Internat. Joint Conf. Pattern Recognition, pp.579-583. 
Deschenes F., Ziou D. (2000) Detection of line junctions and 
line terminations using curvilinear features. Pattern Recognition 
Letters 21:637-649. 
Gautama S., Borghgraef A. (2003) Detecting change in road 
networks using continuous relaxation labeling. In: Proc. ISPRS 
zn 
Workshop "High Resolution Mapping from Space 2003". 
Hannover. 6 pag. 
Haralick R.M., Watson L. (1981). A facet model for image 
data. Computer Graphics Image Proccesing, 15, pp.113-129. 
Levine M., Shaheen S. (1981) A modular computer vision 
system for image segmentations, IEEE Trans. Pattern Anal. 
Mach. Intell. 3(5):540-554. 
Steger C. (1998). An Unbiased Detector of Curvilinear 
Structures. IEEE Trans. Pattern Anal. Mach. Intell. 20(2): 113- 
125. 
Wiedemann C. (2002) Improvement of Road Crossing 
Extraction and External Evaluation of the Extraction Results. 
In: Proc. ISPRS Symposium "Photogrammetric Computer 
Vision" (PCV'02), Graz 2002, Intl. Arch. Photogrammetry and 
Remote Sensing, Vol. 34, Part 3-B, pp. 297-300. -