Page 60 - Dynamic Vision for Perception and Control of Motion

44 2 Basic Relations: Image Sequences – “the World”

about “the world” is affixed to object and subject classes. Of course, the ones of
most importance are those one has to deal with most frequently in everyday life.
Therefore, developing the sense of vision for road or air vehicles requires knowl-
edge about objects and subjects encountered in these contexts most frequently; but
also critical events which may put the achievement of mission goals at risk have to
be known, even when their appearance is rather rare.

2.2.1 Generic 4-D Object Classes
The efficiency of the 4-D approach to dynamic vision is achieved by associating
background knowledge about classes of objects and their behavioral capabilities
with measurement data input. This knowledge is available in generic form, that is,
structural information typical for object classes is fixed while specific parameters
in the models have to be adapted to the special case at hand. Motion descriptions
for the center of gravity (the translational trajectory of the cg in space) forming the
so-called “where”-problem, are separated from shape descriptions, called the
“what”-problem. Typically, summing and averaging of feature positions is needed
to solve the “where”-problem while differencing of feature positions contributes to
solving the “what”-problem. In the approach chosen, the “where”-problem con-
sists of finding the translational transformation parameters in the homogeneous
transformations involved. The “what”-problem consists of finding an appropriate
generic shape model and the best fitting shape and photometric parameters for this
model after perspective projection (possibly including rotational degrees of free-
dom to account for the effects of aspect conditions).
As in computer graphics, all shape description is done in object-centered coor-
dinates in 3-D, if possible, to take full advantage of a decoupled motion description
relative to other objects.

2.2.2 Stationary Objects, Buildings

In road traffic, the road network and vegetation as well as buildings near the road
are the stationary objects of most interest; roads will be discussed in Chapters 7 to
10. Highly visible large structures may be used as landmarks for orientation. The
methods for shape representation are the same as for mobile objects; they do not
need a motion description, however. These techniques are well known from com-
puter graphics and are not treated here.

2.2.3 Mobile Objects in General

In this introductory chapter, only the basic ideas for object representation in the 4-
D approach will be discussed. Detailed treatment of several object classes is done
in connection with the application domain (Chapter 14). As far as possible, the
methods used are taken from computer graphics to tap the large experience accu-
mulated in that area. The major difference is that in computer vision, the actual