= 
landmarkmap | 
  
    
  
   
  
    
  
  
j| LandM 1 
*| LandM2 - 
LandM 3 
AT LandM 4 
ERA] LandM 5 | 
  
  
  
  
  
Geometric. 
Mode 
  
  
   
   
  
+ 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
+ y (k) 
local 
environment y(k) x 
Camera 
Image E 
[A Pee 
BVV2 
| Q estimated LM-Pixel 
-F measured LM- Pixel 
  
estimated. 
cart-position æ 
   
together with the road 
width the curvature para- 
meters yield a very com- 
pact spatial shape 
representation of the road 
in terms of differential 
geometry  [Dickmanns, 
Mysliwetz 92]. The loca- 
tion of the road boundaries 
in the image is mainly de- 
termined by the state vari- 
recursive 
state estimation 
  
  
  
    
  
measured 
cart-position 
  
ATHENE 
Fig.4: Landmark navigation 
ments. Rapid turning of the cart will produce errors partly 
due tothe unpredictable slippage of the wheels. In this case 
the gyroscope carries the best information about the turn- 
ing rate, while the camera is too slow to track detected 
features. The drift of lowcost gyros prevents using the 
signal for a longer period than about one minute. These 
measurement errors and perturbations create a discre- 
pancy between the planned and the real location. There- 
fore depending on the navigational precision required, the 
position of the robot has to be updated by visual feedback 
[Hock, 91](see fig .4). For this purpose there are two 
categories of visual aids to navigation. It will be shown, 
that the combination of both will yield a powerful and 
stable method for traveling autonomously from point to 
point. The first mode is called ' path or lane following" and 
has been well proven over a long period of time on the 
testbed "VaMoRs'. This approach is tailored to well struc- 
tured environments like hallways or paved roads with or 
without lane markings. In path following, motion control 
by visual feedback is limited to one dimension, the lateral 
deviation from the nominal trajectory. Longitudinal con- 
trol only affects time but not the spatial trajectory shape. 
While driving on aroad, the temporal curvature changes 
in a certain look ahead range in front of the vehicle create 
a time varying guidance input to the control system. For 
road image sequence interpretation the assumption is 
made that any change in slope of the road boundaries in 
the image originates either from motion of the vehicle 
relative to the road, from road width changes, or from 
changes in its horizontal and/or vertical direction. Intro- 
ducing road curvature as a state variable to be estimated 
by a Kalman filter was proposed and realized in [Dick- 
manns, Zapp 86] in combination with a dynamical model 
of vehicle motion. As high speed roads exhibit linear 
changes of curvature over runlength, due to ego-motion 
the relations between the curvature parameters can be 
formulated as a compact system of difference equations 
for sampled data systems. Thus, a dynamical model for 
these road parameters also exists. Besides being essential 
for high speed lateral and longitudinal vehicle control, 
  
ables: lateral offset of the 
vehicle from the lane cen- 
ter, camera heading Wg 
relative to the road direc- 
tion and the horizontal and 
vertical road curvature 
parameters. 
There is a difference between autonomous navigation 
in hallways and on roads. Hallways are a guidance net- 
work with predominantly straight connection lines. The 
surface to be travelled on can be considered as a flat and 
smooth plane. Therefore, the lateral deviation from the 
nominal path can be expressed in terms of lateral distances 
to adjacent walls. Any other object with known parameters 
for its geometrical description (environmental model) may 
serve for this purpose as well. The relative lateral distance 
to the object y, and the runlength coordinate then consti- 
tute the state variables. Vehicle displacements from the 
preplanned trajectory may be caused by misalignments 
and odometric errors. 
It is clearly seen that the application of following an 
indoor corridor is a subset of the more complex road 
following task. But as soon as the lane markings disappear 
or a decision has to be made with respect to which object 
or landmark heading has to be selected, an areal navigation 
method is required. The vehicle now has to travel across 
open areas or through extended halls utilizing information 
derived from nearby landmarks. 
The main difference between path and areal navigation 
is that in the first case only topological information may 
be needed, whereas for the two-dimensional case the 
geometrical relations between landmarks must be known. 
The solution for advancing from one landmark to the 
next one will be chosen according to the navigation 
method seemingly preferred by living beings. In this ap- 
proach the absolute distance between two landmarks need 
not be known. Therefore, odometry looses importance in 
this case. This is similiar to the situation when a person 
gives advice on how to reach a certain street intersection. 
Explicit information about the distance to the intersection 
is not needed if the ability of landmark recognition can be 
presumed. A strategy of getting close to the trigger event 
must be known, like following the current street. The 
missing distance information will be substituted by recog- 
nizable patterns, as soon as they get in sight. Since it cannot 
be guaranteed that reliable optical information will be 
Guldance & 
Control