104
ATTENTION CONTROL INTEGRATED IN A SYSTEM TO
ACTIVE TRAFFIC SCENE ANALYSIS
Dirk Wetzel and Karin Sobottka
Bavarian Research Center For Knowledge Based Systems (FORWISS)
Knowledge Processing Research Group
Am Weichselgarten 7, D-91058 Erlangen, Germany
Phone: +49-9131-691187 Fax: +49-9131-691185
e-mail: nameQforwiss.uni-erlangen.de
KEY WORDS: Attention control, traffic scene analysis, motion analysis
ABSTRACT:
' The computation of a robust scene description in real time is still a problem. In analogous to the visual system in
humans a solution can be yield by the integration of attention mechanisms. So an efficient analysis is supported
by focussing the limited resources of a system to those parts of an image that are relevant to the task.
Within this framework an attention control embedded in a system for autonomous driving and collision avoidance
(MOSAIK) is presented. The system is based on the idea to partition the image dynamically in parts for
supervising attention control, detailed recognition, fast tracking and simultaneous recognition and tracking. We
show that the interaction between these processes supports a very efficient analysis also on complex scenes.
The aim of the attention control is alerting the scene for new objects. This is done based on motion analysis
and perceptual grouping strategies.
1. INTRODUCTION
Today computer vision systems in a complex, dynamic world are confronted with a vast flow of information. If
the application requires further a processing in real time, the problem arise to receive robust results in a very
short computing time.
In the case of vision based driver assistance a system have to process about 25 images per second. This video rate
is necessary, because traffic situations may change very quick and decisions have to be as certain as possible.
Existing approaches for traffic scene analysis are often developed as pure tracking components (Dickmanns,
1992) or using special hardware (Masaki, 1992).
An efficient analysis can also be obtained by selecting the interesting parts of each image and focus an detailed
recognition to them (Burt, 1988). This kind of attention mechanism is often developed as single component
(e.g. (Culhane et al., 1992), (Roberts et al., 1993)).
In this framework we present an approach to traffic scene analysis (MOSAIK) (Wetzel, 1995) under egomotion
that interacts between supervising attention control, detailed recognition and fast tracking of objects. This
mechanism makes it possible to compute a robust scene description in nearly real time.
Thereby MOSAIK works fully automatic on monocular image sequences recorded by a graylevel on-board
camera with a resolution of 702 x 566 pixels.
In the following section we give a short overview about the system design of MOSAIK. This will be illustrated
by a so-called cognitive model. Section 3 explains the interaction between attention control, recognition and
tracking. The detection of attention fields is described in section 4. Results are shown in section 5. The paper
concludes with remarks on our future work.
2. THE COGNITIVE MODEL
In MOSAIK each image part is processed as precise and intensive as necessary. Therefore the system interacts
between supervising attention control (A), detailed recognition (R) and fast tracking (T).
In the beginning of the analysis each image will be preprocessed. In this step effects of interlace are removed and
a contrast optimization is done. Then edge information is extracted. Therefore, we apply a modified version
of the filter of Shen and Castan (Castan et al., 1990). Optimal results are received under consideration of two
filter directions (horizontal and vertical). This step is very important to MOSAIK because all system parts (R,
T, A) need edge information as input.
In case of object recognition (Fig. 1, (1)) we apply a step of contour closing to the edge image. So a segmentation
of the image is obtained (Wetzel et al., 1993). Because a robust approach can not assume a perfect segmentation,
the contours of the detected regions are partitioned in straight line segments. Based on this information
geometrical hypotheses of vehicles are generated. A verification of competing vehicle instances is done by
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences’, Zurich, March 22-24 1995