212
5. CONCLUSION
Quantitative and qualitative results of object tracking on the ground plane under egomotion of the camera have been
presented. By applying a single low cost camera, located at the driver's height, high quality position and velocity
estimations have been achieved for objects within the depth of up to 60 m or 20 m respectively. The ratio of the
egocamera height over the road to the depth of the truck is very small, hence the measurement errors are growing
fastly with increased depth. There are two obvious ways to increase the measurement quality for distant objects.
Either the same camera should be located higher over the road or a camera with great focal length should be used.
It is also evident that the quality of velocity estimation could be increased while working with short-term averaged
velocity measurements instead of velocities between two consecutive frames (i.e. synthetic velocity measurement is
done from a short-term tracking of position and direction in 3-5 frames instead of 2).
Acknowledgements: The support from the "Deutsche Forschungsgemeinschaft", Bonn, Germany, is gratefully
acknowledged (Grant Ni-191/8-2). The real images are by courtesy of the BMW AG., Munich, Germany.
REFERENCES
[Kasprzak, 1993] Kasprzak, W. (1993). Modellunabhängige Schätzung von 3-D Attributen während der Bildfol-
gensegmentierung. In: Mustererkennung 1993, Informatik aktuell, 51-58. Springer, Berlin etc.
[Kasprzak et al., 1994] Kasprzak, W., Niemann, H., and Wetzel, D. (1994). Adaptive estimation procedures for
dynamic road scene analysis. In: Proceedings ICIP-94, IEEE Int. Conference on Image Processing, 563-567,
IEEE Computer Society, Los Alamitos, CA.
[Koller et al., 1993] Koller, W., Daniilidis, K., and Nagel, H.-H. (1993). Model-based object tracking in monocular
image sequences of road traffic scenes. International Journal of Computer Vision, 10(3):257-281.
[Masaki, 1992] Masaki, |. (1992). Vision-based Vehicle Guidance. Springer, New York etc.
[Regensburger & Graefe, 1994] Regensburger, U. and Graefe, V. (1994). Visual recognition of obstacles on roads.
In: IROS '94. Proceedings of the IEEE/RSJ/GI International Conference on Intelligent Robots and Systems,
Munich, Germany, 980-987.
[Tan et al., 1993] Tan, T., Sullivan, G., and Baker, K. (1993). Recognizing objects on the ground plane. Image and
Computer Vision, 12(3):164-172.
[Wünsche, 1988] Wünsche, H.-J. (1988). Bewegungssteuerung durch Rechnersehen. Springer, Berlin.
(a) (b) (c)
Figure 4: Vehicles and their hypotheses: (a) left car; (b) truck; (c) badly detectable car.
20 T x : : : jo
Car depth estimation ———
Car depth measurement -------
Truck depth estimation ———
Truck depth measurement --------
PZ [m]
40 +
35 }
30 +
10 1 a 1 n i 1 25 i 1 i i 1 à
Oo 20 40 60 80 100 120 oO 20 40 60 80 100 120
Image Image
Figure 5: The measured and estimated depths (PZ) of the left car (left drawing) and the truck (right drawing).
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
