ent 
ver 
CSS 
“is 
ust 
38, 
ed 
nd 
de 
SS- 
nd 
his 
he 
en 
he 
Human drivers may be very good at the control level 
even in completely unknown environments exploiting 
general knowledge about roads and driveways, traffic 
rules and traffic participants. With respectto guidance, this 
is considered a minor problem assuming some basic navi- 
gational skills and some local support by knowledgeable 
people or correct maps. À similar approach to the overall 
problem of performing a mission has been taken for the 
autonomous computer-guided vehicles. Safe behavioral 
competences in driving have been developed first: the 
necessary capabilities for visual landmark recognition and 
mission performance are added now. 
Intelligent motion control 
The 4D vision process yields the full spatio-temporal 
state of objects including the spatial velocity components 
between objects, if properly set up [Dickmanns, Christi- 
ans 89]. For example, in the road vehicle navigation prob- 
lem both the road curvature parameters in the look-ahead 
range and the state of the own vehicle relative to the road 
may be estimated. With this knowledge a state feedback 
contro] law can be applied in order to obtain a lane 
following competence of the autonomous vehicle[Dick- 
manns, Zapp 86,87; Zapp 88]. 
In order to make the different options (and maybe 
developmental steps) in the evolution of intelligent visual 
road vehicle guidance more clearly visible, several stages 
of control realizations will be discussed. 
Output feedback 
The simplest case is a single control lateral guidance by 
proportional output-feedback to the steering wheel. The 
measured output variable is for example the position of the 
dark to bright transition indicating the road boundary in 
one or several lines of a TV-image of a camera looking 
approximately tangential to the road. If the feedback 
coefficients are properly chosen (and probably adjusted to 
the vehicle speed) already good lateral guidance in simple 
situations can be achieved [Zimdahl et al. 86]. 
Figure 11a gives a block-diagram description of this 
bottom line visual control mode.There is no internal rep- 
resentation of spatio-temporal objects; control is actuated 
in such a manner as to keep the measured value close to a 
predetermined desired one: in our case the position of the 
measured image feature close to a position fixed by the 
designer. 
Implicit notion of state 
In the next step towards intelligence an implicitly avail- 
able model of the process under control allows much more 
flexible control computation. The measured data are 
checked against predicted values derived from internal 
spatio-temporal models.This allows 
1. the elimination of outliers and 
2. intelligent data smoothing exploiting known noise 
statistics. 
    
   
   
   
   
   
   
   
   
   
   
    
    
    
    
   
    
    
  
     
    
   
   
   
   
   
   
    
   
    
   
  
   
   
   
   
     
   
  
    
     
At the core of these recursive estimation methods are 
dynamical models of the process under control capturing 
their typical behavior over time. The essential internal 
variables are directly geared to the physical process in the 
real world, i.e. its state. The measured output variables are 
thus transformed by the estimation process into state vari- 
ables of objects. From this notion, exploiting general 
knowledge from systems dynamics, (optimal) feedback 
laws can easily be derived. Fig.11b shows the correspond- 
ing block diagram. 
The control computation in this mode is still rather fast, 
although much more involved than in case 12a. Because 
of the internal representation of the full physical state of 
the vehicle relative to the road, longitudinal and lateral 
control can be handled easily; even the cross-influence 
from road curvature on acceptable longitudinal speed is 
readily taken care of [Dickmanns, Zapp 86,87]. With to- 
day's microprocessors update rates of 25 Hz are easily 
obtained using modest parallelisation. 
This is a conventional control application not requiring 
any special intelligence(except for the recognition of the 
object road). The performance achieved by this reflex-like 
mode of operation is surprisingly high. Since the underly- 
ing model captures all the essential aspects of the real 
world process rather well, a large variety of lane following 
situations can be handled with just one (possibly adaptive) 
feedback law. In this way, a behavioral competence is 
realized through this special data feedback structure. 
In refined versions of this scheme, the differentiation of 
the internal representation into a situation involving 
several independently represented objects is of advantage 
for more transparency and for obtaining an easier to handle 
interface to the human user. For example, in [Dick- 
manns 88] the road in the look-ahead range and the own 
vehicle have been completely separated yielding a nice 
modular structure; however, all of this is completely im- 
plicit to the program. Up to this point it is just a con- 
venience for the user of the program. In the next step, this 
is going to be exploited for improving behavioral com- 
petences. 
Implicit notion of situations 
Once the notion of spatio-temporal objects and their 
state is available, classes of relative states among objects 
can be recognized requiring similar behavioral actions.For 
example, if the road is free of obstacles, the lane following 
mode with automatic speed adjustment to curvature may 
be run. If, however, an obstacle is encountered, either the 
vehicle has to stop in front of it or it may pass the obstacle 
if there is enough free space to one side. Figure 11c shows 
the program structure for realizing this more flexible type 
of behavior: In parallel to the state estimation it is 
tested,whether there is a candidate for an obstacle on the 
road. If this is true, a sequence of actions may be triggered: 
The longitudinal control mode running is interrupted (the 
lateral one remaining unchanged for the time being) and 
the vehicle is put into a deceleration mode either by