lon
ant
ate
ual
lon
| tO
ing
‘he
sep
wn
her
Che
ep,
ion
2
ion
ted
Handmann, Uwe
according to their capabilities. The processing can be performed for each sensor, also information from different sensors
may be fused as shown by (Handmann et al., 1998c). Objects are extracted by segmentation, classification and tracking
(fig. 3).
Figure 3: Vision-based object detection, object classification and object tracking.
3.0 Sensor-based Representations
The results of the sensor information processing stage are stabilized in movement-sensitive representations by introducing
the time dimension, which was presented by (Handmann et al., 1999). In this sense, a ROI is accepted as a valid hypothesis
only if it has a consistent history. This is implemented by spatio-temporal accumulation using different representations
with predefined sensitivities. The sensitivities are functions of the objects' supposed relative velocity and of the distance
to the observer (fig. 4). In order to apply a time stabilization to these regions and to decide whether they are valid or
not, a prediction of their position in the knowledge integration part is realized. A competition between the different
Figure 4: Image and representation. The value in the representation indicates the activation for an ROI.
representations and a winner-takes-all mechanism ensures reliable object detection. An implementation of an object-
related analysis on vision data has been presented by (Handmann et al., 1998a, Handmann et al., 20002). The results are
passed to the scene interpretation.
4 KNOWLEDGE BASE
The knowledge needed for the evaluation of the data and for information management is determined by the specific task of
driver assistance, physical laws and traffic rules. An improvement of the results can be achieved by the information from
the knowledge base. In the knowledge base, static and dynamic knowledge is represented. Static knowledge is known in
advance independently of the scenery of movement (e.g., physical laws, traffic rules). Dynamic knowledge (e.g., actual
traffic situation, scenery, lane-information) is knowledge changing with the actual information or with the task to be
performed (e.g., objects in front of the car). It can also be influenced by external knowledge like GPS-information. The
accumulation results of the movement-sensitive representations in the object-related analysis, i.e., can be additionally
supported by a preactivation depending on the dynamic knowledge (lane-information, fig. 5).
Figure 5: Oncoming traffic (image with lane-information, movement-sensitive representation without and with preactiva-
tion). The preactivated region is determined by the lane information. The value indicates the grade of activation.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 349