SK = R(So — So,) R(SR — Sn)
R(SE — Se.) R(SP — Sp.)
(1)
with
Sz (5 y, Z,y, 6) (2)
Cosy sing 0) (cos@ 0 —sino
Rz1l-sny cosy 0}-] 0. 1 0 (3)
0 0.1 sin@ 0 cosO
Second an internal model of the object to be tracked is
needed including a rough description of the object
shape and the dynamics of object motion. The dynami-
cal model of the vehicle is approximated by assuming
constant velocity for the object motion parameters. The
object shape is modeled as proposed in [Schick 92] by
using a polygone model with 12 planes, 26 edges and 48
nodes describing the objects surface (figure 4). This
generic shape model is independent of the aspect angle
ofthe viewer and allows a flexible modeling by changing
the form parameters. Because of the symmetry the num-
ber of independent parameters can be reduced to 12.
Figure 4
Generic 3D shape model [Schick 92]
By utilizing these variable parameters nearly every kind
of car type, like limousines, coupes, pickups or trucks
can be coarsely modeled by introducing different metric
proportions for each parameter. For the first hypothesis
of an object it is often sufficient to apply a simple 3D-
rectangular parallelepiped model enveloping the
tracked object. This can be modeled by restricting some
form parameters of the generic shape model (figure 5).
Thus the generic model is easy to adjust to different car
types which are analysed by a separate module for shape
estimation.
/ 8
Figure 5 — Simplified shape model for a truck
IV. HYPOTHESIS GENERATION
Usually, the object detection module searches for ob-
stacles in a certain area of the image in front of the ego
cari depending on the actual curvature of the road,
which is estimated and described by the road module.
This technique of initializing the object tracker works
only for the detection of single and not occluded ob-
jects. In the case of occluded moving objects a more
sophisticated method for generating an initial hypothe-
sis of an object is required. Figure 6 demonstrates the
main components of the extension of the object recog-
nition module for handling situations with occluding
objects. A hypothesis consists of the supposed shape of
the object and an assumed location in the scene. Rec-
ognizing only part of shape of an object in the scene may
produce a valid hypothesis, because we assume that the
tracked object 'Ot" occludes the other region of the
second object.
Objektmanager
Parameter
hypothesis analysis
verification State
estimation Knowledge
base
Feature
hypothesis matching
generation Feature extraction
Video data
Figure6 Components for recognizing occluded objects
Usually, occluded objects appear in front of the tracked
object ’Ot’ at its left or right side. Therefore, the ex-
tended object recognition module for occlusions
searches near the left and the right boundary of object
'Ot for some characteristic features like corner points
or edges, which can be grouped to a partially occluded
box. These features are generated by exploiting the
information of the estimated position of object 'Ot'and
the estimated road curvature. Thus the aspect condi-
tions of the assumed occluded object can be taken into
consideration an analysis can be performed in order to
determine the visible and measureable features, which
are to be extracted from the video image. Matching of
the extracted features to an internal model of the object
is performed by a set of rules from the knowledge base.
Because of the occlusion it is not possible to constrain
the number of features by symmetry. Thus a hypothesis
is generated if a set of features appears which may
belong to an occluded object. In the next section a
method will be discussed for verifying these generated
object hypotheses.
