Page 20 - Dynamic Vision for Perception and Control of Motion

4 1 Introduction

histories have to be observed and a temporally deeper understanding has to be de-
veloped. This is exactly what is captured in the “dynamic models” of systems the-
ory (and what biological systems may store in neuronal delay lines).
Also, through these time histories, the ground is prepared for more compact
“frequency domain” (integral) representations. In the large volume of literature on
linear systems theory, time constants T as the inverse of eigenvalues of first-order
system components, as well as frequency, damping ratio, and relative phase as
characteristic properties of second-order components are well known terms for de-
scribing temporal characteristics of processes, e.g., [Kailath 1980]. In the physio-
logical literature, the term “temporal Gestalt” may even be found [Ruhnau 1994a, b],
indicating that temporal shape may be as important and characteristic as the well
known spatial shape.
Usually, control is considered an output resulting from data analysis to achieve
some goal. In a closed-loop system, where one of its goals is to adapt to new situa-
tions and to act autonomously, control outputs may be interpreted as questions
asked with respect to real-world behavior. Dynamic reactions are now interpreted
to better understand the behavior of a body in various states and under various en-
vironmental conditions. This opens up a new avenue for signal interpretation: be-
side its use for state control, it is now also interpreted for system identification and
modeling, that is, learning about its temporal behavioral characteristics.
In an intelligent autonomous system, this capability of adaptation to new situa-
tions has to be available to reduce dependence on maintenance and adaptation by
human intervention. While this is not yet state of the art in present systems, with
the computing power becoming available in the future, it clearly is within range.
The methods required have been developed in the fields of system identification
and adaptive control.
The sense of vision should yield sufficient information about the near and far-
ther environment to decide when state control is not so important and when more
emphasis may be put on system identification by using special control inputs for
this purpose. This approach also will play a role when it comes to defining the no-
tion of a “self” for the autonomous vehicle.

1.3 Why Perception and Not Just Vision?

Vision does not allow making a well-founded decision on absolute inertial motion
when another object is moving close to the ego-vehicle and no background can be
seen in the field of view (known to be stationary). Inertial sensors like accelerome-
ters and angular rate sensors, on the contrary, yield the corresponding signals for
the body they are mounted on; they do this practically without any delay time and
at high signal rates (up to the kHz range).
Vision needs time for the integration of light intensity in the sensor elements (33
1/3, respectively, 40 ms corresponding to the United States or European standard),
for frame grabbing and communication of the (huge amount of) image data, as well
as for feature extraction, hypothesis generation, and state estimation. Usually, three
to five video cycles, that are 100 to 200 ms, will have passed until a control output