A large fraction of our knowledge about the real world 
is concerned with the temporal domain; we learn to under- 
stand this during early life more or less subconsciously 
while the capability of crawling, walking and manipulat- 
ing other objects under earth gravity is being acquired. The 
temporal sequence of states of moving objects and their 
transition characteristics constitute very essential knowl- 
edge about the real world providing us with the capability 
of acting adequately even though it does not seem to be 
represented explicitly. This has long been overlooked in 
Artificial Intelligence which concentrated its efforts on 
explicitly represented abstract knowledge about quasi- 
static relations between objects in the world. 
The natural sciences and engineering technology have 
developed adequate methods for representing these facts 
about the physical world. They describe them within the 
framework of differential equations with time as mono- 
tonically increasing independeni variable. As I.Kant has 
elaborated in his 'Critiques ...' more than two centuries 
ago it has to be kept in mind that space and time are not 
properties of objects. We cannot help carrying it into the 
world by our sensing and analysis systems; we ourselves 
exist in these basic four dimensions. Therefore, it was 
decided to install these basic four dimensions in the 4D- 
approach to dynamic machine vision right from the begin- 
ning in order to be able to deal with the real world 
efficiently. This was the main contribution of our approach 
to machine vision; the rest follows almost automatically. 
2. System components 
  
The availability of two different testbeds each with a 
navigation system based on the 4D-approach [Dickmanns, 
Graefe, 88], allows test runs to be performed under various 
  
Fig.1: Experimental vehicles: a) VaMoRs b) ATHENE 
   
   
   
   
   
    
   
    
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
    
    
   
   
    
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
    
kinds of circumstances. The first vehicle is clearly 
specified to indoor applications and called 'ATHENE', 
whereas the second one, dubbed ' VaMoRs', is designed 
for outdoor usage. Each system has its specific advan- 
tages, but the overall design of the navigation system is 
closely related, so there are no difficulties to transfer well 
proven solutions between the two. Since ' VaMoRs' is well 
documented in [Zapp 88; Dickmanns, Graefe 88] more 
effort is put on describing details of 'ATHENE' in this 
paper. 
The main components of the indoor experimental setup 
can be divided into three categories. First, the robot itself, 
which is a converted AGV equipped with all necessary 
actuators, interfaces and onboard power supply. Second, 
a special multiprocessor vision system for realtime image 
sequence analysis and interpretation is placed on the ve- 
hicle, and third, there are two extra computers for the 
navigational task and the low level control system. 
ATHENE can be driven autonomously under computer 
control and serves as a rolling indoor platform for research 
work on landmark navigation and computer vision. 
In the front part of the vehicle an electromechanical pan 
platform is mounted carrying a standard monochrome 
CCD camera. Viewing direction control is done either by 
the navigation system or the vision system itself, depend- 
ing on the actual task. The camera pointing capability 
allows active scene search and horizontal tracking, e.g. for 
initial self orientation or landmark tracking while driving. 
For image sequence processing a custom made system 
BVV 2 [Graefe 90] with four Intel 80286 processors has 
been utilized in the experiments. This multiprocessor sys- 
tem of the MIMD type consists (in the case of the indoor 
application) of 4 commercial, standard Multibus I single- 
board computers spanning the performance range from 
8086 to 80286. Key feature of this multiprocessor vision- 
system is the physically distributed, thus truly parallel 
image access capability of all CPUs directly involved in 
image operations. This overcomes the common I/O bot- 
tleneck of general purpose machines, in which usually 
only one processor has direct image access. Here, no 
central frame store exists. Any processor linked to the 
videobus through a custom made videobus-interface 
(VBI) can simultaneously access and process a subseg- 
ment (window) of the digitized 256 by 244x8 bit per pel 
grayscale image. The VBI basically is a hardware-attach- 
ment to a standard singleboard computer containing a 
window-selection logic and two fast window-buffers stor- 
ing 4k pel each. Multiple windows can be independently 
positioned or changed in size, shape and sampling density 
under software control. It should be noted that except for 
the VBI no custom hardware and no dedicated image 
processing devices are being used in this experimental 
system. The advantages of applying easily programmable 
standard microprocessors instead, proved to be significant 
for the system's applicability and efficiency as a research 
tool. A further key point is the flexible interprocessor 
communication scheme based on message passing, form-