Page 144 - Dynamic Vision for Perception and Control of Motion

128 5 Extraction of Visual Features

stripes, whose orientation and width have to be chosen intelligently ad hoc, exploit-
ing known continuity conditions in space and time. These stripes will be condensed
to one-dimensional (averaged over the width of the stripes) representation vectors
by choosing proper schemes of symbolic descriptions for groups of pixel values in
the stripe direction.
Coming back to the difference between data and information, in knowledge-
based systems, there may be a large amount of information in relatively few data, if
these data allow a unique retrieval access to a knowledge base containing informa-
tion on the object recognized. This, in turn, may allow much more efficient recog-
nition and visual tracking of objects by attention focusing over time and in image
regions of special interest (window concept).
According to these considerations, the rest of Chapter 5 will be organized as fol-
lows: In Section 5.1.2, proper scaling of fields of view in multi-focal vision and in
selecting scales for templates is discussed. Section 5.2 deals with an efficient basic
edge feature extraction operator optimized for real-time image sequence under-
standing. In Sections 5.3, (2-D) region-based image evaluation is approached as a
sequence of one-dimensional image stripe evaluations with transition to symbolic
representations for alleviating data fusion between neighboring stripes and for im-
age interpretation. An efficient method with some characteristics of both previous
approaches is explained in Section 5.4; it represents a trade-off between accuracy
achievable in perception and computational expense.
Contrasting the feature extraction methods oriented towards single-object rec-
ognition, Section 5.5 gives an outlook on methods and characteristic descriptors for
recognizing general outdoor environments and situations. Computing power avail-
able in the past has not allowed applying this in real-time onboard vehicles; the
next decade should allow tackling this task for better and more robust scene under-
standing.

5.1.2 Fields of View, Multi-focal Vision, and Scales

In dealing with real-world tasks of surveillance and motion control very often cov-
erage of the environment with a large field of view is needed only nearby. For a
vehicle driving at finite speed, only objects within a small distal range will be of
interest for collision avoidance. When one is traveling at high speed, other low-
speed objects become of interest for collision avoidance only in a rather small an-
gular region around the subject’s velocity vector. [For several high-speed vehicles
interacting in the same space, special rules have to be established to handle the
situations, such as in air traffic control, where different altitudes are assigned to
airplanes depending on their heading angle (in discrete form by quadrants).]
Central projection is the basic physical process of imaging; depending on the
distal range of the object mapped into the image, this results in one pixel represent-
ing areas of different size on objects in the real world. Requiring a certain resolu-
tion normal to the optical axis for objects in the real world, therefore, requires
range-dependent focal lengths for the imaging system.
Biological systems have mastered this problem by providing different pixel and
receptive field sizes in the sensor hardware eye. In the foveal area designed for