mation, a scheme
; proposed. The
3, assuming the
image, perform
. the constrained
(Kanade et al.,
e and the current
next right image
tively (the arrow
n figure 3).
t right image are
is determined. In
a matching point
A“ and the point
eppipolar
line
| constrained
matching
range
e, and the above
Finally, using
ng 3D point of a
will act as an
OACH
trade-off can be
putting from the
trajectory is low,
1=0.5. To keep a
re used. @;=0.5,
enting historical
time-consuming
sion in equation
time. In our
] every 2.0-2.5
the inversion is
10-20 according to the vehicle velocity and image capture rate.
Without using B-splines curve representation, the dimension
(n) would be 80-160 (assuming sampling interval on the curve
is 0.3 meters); (b) Due to the known orientation parameters of
cameras and a 3D road shape model, the processing window
constraint for feature extraction and the constrained matching
range for feature matching are available. The computation
problem of the highly time-consuming part is greatly reduced.
5.2 Test Results
The software based on this approach has been implemented
and installed in the ImagExpert system which is a software
package for automatic processing of VISAT™ images. This
package 1s developed on a SUN SPARC workstation with an
X-Window and MOTIF environment. A high level object-
oriented development toolkit WNDX was used to design the
Graphic User Interface. The package has been also interfaced
with AUTOCAD V12. The built-in communications between
our package and the AUTOCAD core was implemented. The
process of the reconstruction can be displayed in AUTOCAD
environment simultaneously.
The present approach has been widely tested on real image
sequences captured by VISAT™ system. Tests have been
conducted to evaluate the approach with images under different
road conditions, such as roads with missing, intermittent and
faded centerline markings caused by shadows, cracks and wet
roads, solid and dashed line patterns, and sharp curves and
gaps. The typical examples are given in figure 4-6 (dotted
black lines represent the results). On these complex cases, the
approach performed successfully and reliably. Guided by our
road model, no failure problems occurred even in big gaps and
turns. The accuracy of the reconstruction result will be
discussed in a separate paper.
6. CONCLUDING REMARKS AND
FUTURE WORK
An integrated approach has been proposed to synthesize
multiple constraints from image sequences, in which the
reconstruction of road centerlines has been considered as a
"shape form sequences" problem. Based on our tests, the
approach has the following characteristics:
e Physically-based mechanism: The synthesis of multiple
constraints has been implemented by the least action principle,
a generalized principle of physical motion. The advantage of
the method is that constraints both from object assumptions
and image sequences can be incorporated into a model.
e Model initialization: The highly accurate vehicle
trajectory determined by a multi-sensor (GPS/INS) integration
technique is used for generating an approximate 3D shape
model of road centerlines.
e Model based reconstruction: The whole process of the
reconstruction is designed based on a 3D shape model. This
design leads to a robust result because the shape model is
deformed incrementally under the combinations of the actions
of the internal and external energy.
e Model driven feature extraction and matching: Effective
geometric constraints are available for feature extraction and
matching on images with large geometric discrepancies. The
combined processing of stereo and motion image can also be
applied.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
e B-Splines modeling (3D B-snake): Not only geometric
characteristics, but also numeric advantages can be achieved in
3D curve reconstruction.
Due to the robustness of the approach, in future work we
intend to combine this approach with a Kalman filter based
navigation algorithm. In this sense, accurate navigation using
multiple sensors and 3D line feature reconstruction are
integrated and solved in a combined case. This scheme is used
to take both navigation error and reconstruction error into
account and to further improve the global accuracy of both
results.
7. ACKNOWLEDGMENTS
I would like to give special thanks to Dr. R. Li and Dr. M. A.
Chapman who introduced me to this research and gave me
invaluable comments. I would like to thank Dr. K. P. Schwarz
who led the project from an original idea to a production
system. I would also like to thank my colleagues, Mr. L. Qian
and Mr. Y. Xu for their great help on implementing the
system. Mr. Derek Lichti’s help on language correction is
appreciated. Finally, the support of Canadian Natual Science
and Engineering Research Council (NSERC) and GEOFIT Inc.,
Laval, Quebec is appreciated.
8. REFERENCES
H. Baker, R. Bolles, and J. Woodfill, 1994. Real Stereo and Motion
Integration For Navigation, Int Archives of ISPRS Com. III symposium.
Munich, SPIE-2357, pp. 17-24.
N. El-Shemy, and K. P. Schwarz, 1993. Kinematic Positioning in
Three Dimensions Using CCD Technology. Vehicle Navigation and
Information Sysiem'93 Conf., Ottawa.
G. He, and K. Novak, 1992. Automatic Analysis of Highway
Features from Digital Stereo Images. Int. Archives of ISPRS Com. III, part
B3, pp.119-124.
T. Kanade, M. Okutomi, and T. Nakahara, 1992. A Multiple-
baseline Stereo Method, Proc. of ARPRA Image understanding Workshop.
San Diego, CA, pp. 409-426.
M. Kass, A. Witkin, and D. Terzopoulos, 1988. Snakes: Active
Contour Models, Int. J. of Comp. Vision, pp. 321-331.
R. Li, K. P. Schwarz, M. A. Chapman, and G. Marcel, 1994.
Integrated GPS and Related Technologies for Rapid Data Acquisition,
GIS World, Vol.7, No.4, 41-43.
S. Menet, P. Saint-Marc, and G. Medioni, 1990. Active contour
models: overview, implementation and applications, Int. Conf. Syst. Man
Cybernet. pp. 194-199, Los Angeles.
H. Schneiderman, and M. Nashman, 1994. A Discriminating
Feature Tracker for Vision-Based Autonomous Driving, IEEE Trans. on
Robotics and Automation. Vol. 10, No. 6, pp. 769-775.
K. P. Schwarz, H. Martell, N. El-Sheimy, R. Li, M. Chapman, and
D. Cosandier, 1993. VISAT - A Mobile Highway Survey System of High
Accuracy, Vehicle Navigation and Information System conference'93,
Ottawa, October 12-15.
C. Tao, 1996. Road Centerline Reconstruction from Seugential
Images Based on Shape from Sequences, SPIE proceedings of Visual
Communciations and Image Processing’96, Orlando, Florida, USA,
March 17-20.
C. Tao, R. Li, and M. A. Chapman. 1996. A Model Driven
Approach for Extraction of Road Line Feature Using Stereo Image
Sequences From A Mobile Mapping System, to be published in the
proceedings of ASPRS/ACSM Annual Convention, April. Baltimore,
Maryland.
D. Terzopoulos, A. Witkin, and M. Kass, 1988. Constrains on
deformable models: Recovering 3D shape and nonrigid motion, Artificial
Intelligence, Vol. 36, pp. 91-123.
C. Thorpe, et al, 1988. Vision and Navigation for the Canegie-
Mellon Navlab, IEEE Trans. on PAMI, Vol. 10, No. 3, pp. 362-372.
