00 "% pairs 
Putative ones 
ulate (Eq. 3) 
le d by calculating 
llert «.al, 1%7) 
re 50.50 m, 64.02 
by incorrect 
ite the accuracy of 
«some data pre . 
of outliers. To 
ition rule based on 
r than mean and 
is called “Hubber 
og rule: Reject 
rations away from 
der (Hampeit al., 
absolute values of 
gure 7). In a set of 
eme outliers that 
f such errors the 
always very large. 
jrs were analysed 
5 considered as a 
n the set of 55 
utliers the average 
apdi 
iverage 
road width 
O.H in. ^ 
neasurement is 
reason 
for 
correspondence fails in range of 5 to 50 pixels (sometimes even 
more). Such a mismatch causes metres error in distance of 15 m 
from the cameras. Mismatch are mainly caused by: 
1. Significant difference in image radiometry. 
2. Perspective effect (foreshortening). 
3. Errors in interior and exterior calibration (weak 
reconstruction of epipolar geometry) 
4. JPEG compression (bad image quality) 
To overcome problems 2 and 3 the system needs to be precisely 
calibrated. Because of vibrations of the camera system, a 
correction to the orientation parameters should be determined 
frequently, e.g., every 300 image pairs. The significant difference 
in image radiometry can be avoided by using an additional 
sensor for measurement of light intensity. The image quality can 
be improved, e.g., by using non-destructive JPEG algorithm and 
by using cameras with the same precise settings of colours. 
4 CONCLUSIONS 
The algorithm for automatic measurement of the road was 
developed and tested on real life images. The measurements of 
the road width from georeferenced images collected by MMS are 
very limited. Accuracy of any point on the road boundary is 
correlated to the distance/base ratio and the distance from the 
road centreline. If precision better than 0.5 m is required, 
measurements of the road width are restricted to narrow roads 
only, where the closest point on the road boundary is within a 
distance shorter than 20 m. Good results can be obtained only if 
the measurements are performed semi-automatically and if the 
road is narrow and have well defined boundaries, e.g., painted 
white lines. In this case the precision of measurements is in range 
of 0.11 m to 0.3 m. 
Developed automatic methods require images with high 
geometric and radiometric quality together with well-determined 
interior and exterior orientation parameters. The segmentation of 
the road is successful even if the road has no well-defined 
boundaries. Unfortunately, due to mismatch (see Section 3) 
precise measurements can not be performed properly. Automatic 
algorithm requires better road model (see Section 2.1) and 
support from other data sources e.g., existing road databases. The 
stereo image analysis should be extended to stereo sequence 
analysis with data association technique and Kalman filtering. 
5 REFERENCES 
Crisman J.D., Thorpe C.E. 1993. SCARF: A Colour Vision 
System that Tracks Roads and Intersection.IEEE Transactions 
on Robotics and Automation. Vol 9. No 1. pp 49 - 58. 
Duda R., Hart P. 1972., Use ofHough Transformation to Detect 
Lines and Curves in a Pictures. Communications of the ACM, 
Vol. 15, No. 1 (January), pp 11-15. 
Gajdamowicz K. 1994. A Mobile Data Collecting System Based 
On GPS and Video Cameras. Trita/Geod-94/3039, Dept, of 
Geodesy and Photogrammetry. KTH, Stockholm, Sweden. 
Gajdamowicz K. 1998. Automated Processing of Colour Stereo 
Georeferenced Images for Road Inventory. International 
Archives of Photogrammetry and Remote Sensing, Vol.32, No.5, 
pp 690-698. 
Habib A.F. 1994. Estimation of Motion Parameters for Stereo- 
Image Sequences Using Data Association of Linear Features. 
Dissertation. Ohio State University. 
Hampel F. R., Ronchetti E.M., Rousseeuw P.J, Stahel W.A., 
1986. Robust Statistic. John Wiley &Sons, Inc. pp 62-71 
Hallert B., Kaasila P., Ottoson L., Öhlin K..A., 1967. Quality 
Problems In Photogrammetry. International Society for 
Photogrammetry. Commission II. Thdntercomissional Working 
Group On Fundamental Problems. Stockholm. 
Kluge K., Thorpe C.E. 1989. Explicit Models for Robot Road 
Following. IEEE International Conference on Robotics and 
Automation, Scottsdale AZ, pp 1148-1154 
Mayer, H., Laptev, L, Baumgartner, A. 1998. Multi-Scale and 
Snakes for Automatic Road Extraction. Fifth European 
Conference on Computer Vision (2), pp 720-733. 
Nalwa V.S., 1993. A guided tour to computer vision Addison- 
Wesly Publishing Company. ISBN 1 -201 -54853-4., pp 200-209. 
Tao Ch. 1996. An Integrated Approach To Road Centreline 
Reconstruction Using Stereo Image Sequences From A Mobile 
mapping System. International Archives of Photogrammetry and 
Remote Sensing, Vol.31, No.B3, pp.857-862. 
Turk M., Morgenthaler D.,Gremban K.,Marra M., 1988. “VITS: 
A Vision System for Autonomous Land Vehicle Navigation. 
IEEE Transaction on Pattern Analysis and Machine Intelligence, 
Vol. 10, No. 3, pp 342-361 
Wallace R.,Matsuzaki K.,Goto Y., Crisman J.,Kanade T. 1986. 
Progress in road following” Proc. IEEE .International Conference 
on Robotics and Automation. San Francisco. 
Yang H., 1995. Vision Methods for Outdoor Robot Navigation. 
PhD Thesis. . ISSN 0783-5477. Helsinki University of 
Technology. Finland. 
7B-2-5