Page 59 - Dynamic Vision for Perception and Control of Motion

2.2 Objects 43

scale if computing power is limited. This approach has proven to be efficient and
robust.
Even with this approach, to cover a range of possible object distances of two to
three orders of magnitude, the size of objects in the images still varies over more
than one order of magnitude; this fact has to be dealt with. Pyramid techniques
[Burt et al. 1981] and multiple scales of feature extraction operators are used to
achieve this. This requires that the same object be represented on different scales
(with different spatial resolution and corresponding shape descriptors). Homoge-
neous coordinates allow representing different scales by just one parameter, the
scaling factor. In the 4 × 4 transformation matrices, it enters at position (4, 4).

2.1.4.2 Multiple Timescales
On the time axis, the lower limit of resolution is considered the cycle time of elec-
í8
tronic devices such as sensors and processors; it is presently in the 10 to 10 í10
second (s) range. Typical process elements in digital computing such as message
overheads for communication with other processing elements last of the order of
magnitude of 0.1 to 1 ms; this also is a typical range of cycle times for conven-
tional sensing as with inertial sensors.
The video cycle times mentioned above are the next characteristic timescale.
Human reaction times are characterized by and may well be the reason for the
1ísecond basic timescale. Therefore, the 0.1 to 10 s scale ranges are the most im-
portant and most frequently used. “Maneuvers” as typical time histories of control
outputs for achieving desired transitions from one regime of steady behavior to an-
other last up to several minutes. Quasi-steady behaviors such as road-running on a
highway or flying across an ocean may last several hours. Beyond this, the astro-
nomically based scales of days and years predominate. One day means one revolu-
tion with respect to the sun around the Earth’s axis; it has subdivisions into 24
hours of 60 minutes of 60 seconds each that is 86 400 seconds in total. One “year”
means one revolution of Earth around the Sun and includes about 365 days. The
corresponding lighting and climatic conditions (seasonal effects) affect the opera-
tion of vehicles in natural and man-made environments to a large degree.
The lifetimes of typical man-made objects are the order of 10 years (vehicles,
sensors); human life expectancy is 5 to 10 decades. Objects encountered in the en-
vironment may be hundreds (trees, buildings, etc.) or many thousands of years of
age (geological formations). Therefore, also in the temporal domain the range of
10
í9
scales of interest spans from 10 to about 10 seconds or 19 orders of magnitude.
Autonomous systems to be developed should have the capability of handling this
range of scales as educated humans are able to do. In a single practical application,
the actual range of interest is much lower, usually.

2.2 Objects

Beside background knowledge of the environmental conditions at some point or
region on the globe and the variations over the seasons, most of our knowledge