In the "Autobahn" driving application the estimated 
motion and location parameters of the object 'Ot and 
the occluded object should be similar, because both of 
them are driving in the same direction on the same road. 
But the position state components may not be identical, 
for this case the two objects would occupy the same 
space. They may be close to each other, e.g. a truck 
linked with its trailer, and the assumed hypothesis of two 
independently moving objects may have been wrong. 
With this method it is also possible to distinguish objects 
standing or parking beside the road, being uncovered, 
When the object 'Ot' continues to move in front of the 
own vehicle. By this way it is verified that the generated 
hypothesis of an occluded object is either a casually 
existing object in the surrounding like bridges or stakes 
nor a part of the originally tracked object 'Ot', because 
of a not exactly matching shape model. Therefore, this 
module is also able to provide information to the situa- 
tion assessment module about objects located near by 
the road boundary. 
In this approach decisions are made by analyzing the 
estimated motion and position parameters statistically 
over a period of time with respect to the existence of a 
second partially occluded object. If the hypothesis of an 
occluded object was verified the algorithm tracks the 
object until it disappears. 
VI. IMPLEMENTATION AND EXPERIMENTAL 
RESULTS 
To verify the practicality of the proposed approach the 
extended object recognition module has been imple- 
mented. All the experimental work is done in a closed 
loop simulation consisting of a graphics workstation for 
image generation and a parallel-processor system for 
image processing. Two possible modes of operation are 
implied. In the first one, only the original synthetic 
images of the workstation are used to test the imple- 
mented algorithms with a cycle time of 80 ms. In the 
second one, a real CCD-camera takes the images from 
the graphic screen, and all the problems using noisy 
measurement data under different viewing and illumi- 
nation conditions were covered. All implementations 
have been done in C to achieve the real-time demands. 
Up to now, all research work was performed in the 
software simulation environment without using real im- 
ages from a CCD-camera. The results achieved show, 
that the analysis of a hypothesis takes about 20 video 
cycles to make a decision about the assumption of an 
partially occluded object. 
The following figures show the estimated state variables 
for the object location (distance, lateral offset) of two 
cars one overtaking the other by using simulated noisy 
measurements. Figure 9 illustrates the results produced 
by the verification algorithm. 
     
   
Cycle 
results from motion analysis 
  
29 generating Hypothesis of an second occluded 
object 
29 :  Initialisation Hypothesis with identical motion 
42 second occluded object moving in front (right) 
108 two different motion types 
137 . changing Hypothesis 
202 second occluded object moving in front (left) 
234 canceling Hypothesis of an occluded object 
Figure 9 Analysis of a motorway situation with an occluded object 
Exploiting the estimated velocities becomes very diffi- 
cult, because the estimation error has the same magni- 
tude as the estimated state variable. Therefore it seems 
to be sufficient to analyse the position parameters. In 
the case, that velocities are to be analysed additionally 
it is recommended to smooth the state variables in a 
succeeding low-pass-filter (figure 13). Figures 10a and 
b show the estimated distance of the two objects. The 
second object hypothesis is initialized at cycle 29 with 
the initial assumption that the occluded object belonges 
to the already tracked object 'Ot'. At cycle 42 an oc- 
cluded object was verified and the motion analysis re- 
sults with the statement that an other object is moving 
Abstand x: xs(rot--], real[blau-], PS Dx[gruen-.] * 10 Y 
  
120} T 7 = 
softs Ri: rit ni deer pp ri, nd 
a ius 
  
  
  
  
kafil 30 Zeh 
Figure 10a Estimated distance of object Ot 
Abstand x: xs[rot--], real[blau-], PS Dx[gruen-.] * 1 
120r dry = : ] 
  
80 r- mei 
i / 
i i a par 
60 fe nm arm ma t ii mon ig SEN mt o 
E si. - 
eS n fT 
  
20r-- i . E CR ea —— E : E 
  
  
  
50 100 150 200 250 
kafi2 30 Zyklen 
Figure 10b Estimated distance of the occluded object