iin
ted
nce
for
ysis
If
ery
ate
le.
ns,
led
ent
ion
his
ard
ted
nd
per
cts
nd
ion
wo
ion
on,
ion
95
105
considering the temporal context and geometrical structure of hypotheses. Hereby a fuzzy description of a
vehicle model is matched with groups of geometrical hypotheses over time (Wetzel et al., 1994). To make a
reaction planning possible, a 4D-description of the scene has to be computed. This is realized by reprojection
of the road line segments from the 2D image in the 3D-world. Therefore a flat road have to be assumed.
planning
m scene- — --....
e = interpretation ur
MT ee "e + Se
T | se . *
* *
° o Se ss
prediction Fr # hypotheses- >, N prediction
^d Á * oque . * .
SL / verification son 2)
J : * .
: ; 3 x
i J 1 *
J : j ; 1 1
i $iwarning RS warning *
i i (6) refreshing, :
: : ' :
À q) o
5 , 1 1
A (3 “©
LÀ hypotheses- 2 +
à A generation ls
V X unknown "i. known 74. .
X objects..." tenes, Objects," e
Weer (7) ( 4) >=. g
attention : tracking
eu 0 0 et 1 vue >
preprocessed "| iet preprocessed
eo US
images
Figure 1: The cognitive model
When objects are reliable recognized, they can be tracked based on graylevel values and edge information. By
predicting the object position in the next frame (Fig. 1, (2)) a fast refreshing of the object attributes (3) (e.g.
motion, position in relation to the road, distance to the camera vehicle) is realized. In case of uncertainty the
corresponding image part of an object turns back to recognition (4).
The attention control is continuously searching for new objects in the scene. Therefore it works only on those
parts of the image that are not processed by the tracking or recognition component of MOSAIK (5). To detect
regions of interest motion is determined and line segments which are moving in a similar way, are grouped
together under consideration of parallelism, orthogonality etc. (see section 5). If regions of interests, so-called
attention fields, are detected, the attention control send a warning (6) and the corresponding image part is
turned back to the recognition in MOSAIK (7).
This cognitive model is represented by a semantic net (Niemann et al., 1990). It consists of 16 concepts, which
are linked by different types of links (Wetzel et al., 1994). In the network 7 levels of abstraction are distinguished.
Intensive recognition is done on the highest level of abstraction. Attention control and tracking work on the
lower abstraction level 2.
Based on the interaction of recognition, tracking and attention control a scene interpretation is determined and
a reaction planning can be done.
3. INTERACTION IN MOSAIK
MOSAIK is based on the idea to partition the image dynamically in parts for detailed recognition, fast tracking,
supervising attention control and simultaneous recognition and tracking (STR). An example of a partition of
an image after the initial phase is shown in the left of figure 2.
In the initial phase of the analysis (Fig. 2 right, (1)), when all objects are unknown, the recognition works on
the whole image. If an object is reliable recognized, it changes to the status of tracking (2). In the case, that
the attributes of a tracked object change in an unpredictable way, simultaneous to the object tracking a renew
recognition is initiated (3). After a successful correction of the object borders, the object changes back in the
status of pure tracking (4). The initial phase is finished, when all objects are in the status of tracking. At this
time the recognition is turned off (5). In the following dynamic phase the attention control alerts the scene for
new objects. If a new object appears, the attention control responds to it and the corresponding image part
changes to the status of recognition (6). There the hypotheses of the attention control will be verified.
If an object leave the visible part of the camera, the image part is at the status of attention control.
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995