Page 131 - Dynamic Vision for Perception and Control of Motion

4.1 Structuring of Application Domains 115

bases and algorithms for recognizing these items; they have to be swapped when
entering a zone with new regulations.
In section Appendix A.3 the different types of vehicles are listed. They have to
be recognized and treated according to their form (shape), appearance and function
of the vehicle (Appendix A.4). This type of structuring may not seem systematic at
first glance. There is, of course, one column like A.4 for each type of vehicle under
A.3. Since this book concentrates on the most common wheeled vehicles (cars and
trucks), only these types are discussed in more detail here. Geometric size and 3-D
shape (Appendix A.4.1) have been treated to some extent in Section 2.2.3 and will
be revisited for recognition in Chapters 7 to 10.
Subpart hierarchies (Appendix A.4.2) are only partially needed for vehicles
driving, but when standing, open doors and hoods may yield quite different ap-
pearances of the same vehicle. The property of glass with respect to mirroring of
light rays has a fundamental effect on features detected in these regions. Driving
through an environment with tall buildings and trees at the side or with branches
partially over the road may lead to strongly varying features on the glass surfaces
of the vehicle, which have nothing to do with the vehicle itself. These regions
should, therefore, be discarded for vehicle recognition, in general. On the other
hand, with low light levels in the environment, the glass surfaces of the lighting
elements on the front and rear of the vehicle (or even highlights on windscreens)
may be the only parts discernible well and moving in conjunction; under these en-
vironmental conditions, these groups are sufficient indication for assuming a vehi-
cle at the location observed.
Variability of image shape over time depending on the 3-D aspect conditions of
the 3-D object “vehicle” (Appendix A.3) is important knowledge for recognizing
and tracking vehicles. When machine vision was started in the second half of the
last century, some researchers called the appearance or disappearance of features
due to self-occlusion a “catastrophic event” because the structure of their (insuffi-
cient) algorithm with fixed feature arrangements changed. In the 4-D approach
where objects and aspect conditions are represented as in reality and where tempo-
ral changes also are systematically represented by motion models, there is nothing
exciting with the appearance of new or disappearance of previously stable features.
It has been found rather early that whenever the aspect conditions bring two fea-
tures close to each other so that they may be confused (wrong feature correspon-
dence), it is better to discard these features altogether and to try to find unambigu-
ous ones [Wünsche 1987]. The recursive estimation process to be discussed in
Chapter 6 will be perturbed by wrong feature correspondence to a larger extent
than by using slightly less well-suited, but unambiguous features. Grouping re-
gimes of aspect conditions with the same highly recognizable set of features into
classes is important knowledge for hypothesis generation and tracking of objects.
When detecting new feature sets in a task domain, it may be necessary to start
more than one object hypothesis for fast recognition of the object observed. Such
4-D object hypotheses allow predicting other features which should be easily visi-
ble; in case they cannot be found in the next few images, the hypothesis can be dis-
carded immediately. An early jump to several 4-D hypotheses thus has advantages
over too many feature combinations before daring an object hypothesis (known as
a combinatorial explosion in the vision literature).