Page 148 - Dynamic Vision for Perception and Control of Motion

132 5 Extraction of Visual Features

the scale range is covered by different mask sizes while the other half is handled on
larger scales by image size reduction.
Literature on edge feature extraction abounds (see, e.g., World Wide Web:
http://iris.usc.edu/Vision-Notes/bibliography/contents.html; detection and analysis
of edges and lines (Chapter 5 there); 2-D feature analysis, extraction, and represen-
tations (Chapter 7); Chapter 3 there gives a survey on books (3.2), collections,
overviews, and surveys).

5.2.1 Generic Types of Edge Extraction Templates

A substantial reason for the efficiency of the methods developed at UniBwM for
edge feature extraction stems from the fact that both low-pass filtering in one di-
rection and accurate search in an almost normal direction were combined. By
proper partitioning of the overall task, simple but efficient pixel processing in just
one dimension has been achieved consecutively. It is assumed that the direction of
the edge to be found is known approximately. While this is true for tracking, it
turned out that the algorithms are also useful for initialization if proper parameter
settings are chosen. The software packages developed for edge extraction are gen-
eralizations of the Prewitt operator [Ballard, Brown 1982]; they have matured over
three generations of coding in different computer languages. The original ideas of
Kuhnert (1988) have been refined and coded first in the second half of the 1980s in
the language FORTRAN by Mysliwetz (1990); the current version was developed in
the early 1990s under the name KRONOS D.Dickmanns (1997) in the language Oc-
cam for transputers. For the next generation of processors, it has been converted to
C under the name CRONOS and polished by S. Fuerst.

5.2.1.1 Low-pass Filtering of an Oriented Pixel Field into a Vector

To obtain good correlation values for a local edge extraction operator, its orienta-
tion should be almost tangential to the edge in the image. Figure 5.6 shows a trape-
zoidal dark area in front of a pixel grid, the edges of which are to be detected. The
mask for edge detection shown to the
left in the pixel grid has approximately Correlation Image edges
the same inclination as the left edge of mask
the area to be detected; the mask is n w =
17 pixels wide and m d = 5 pixels deep. Window
To be independent of absolute light in-
tensity, the number of plus signs and
minus signs in the mask has to be equal;
in the case shown, there are two pixel Search
formations along the edge (called “mask Threshold path
elements” henceforth, one pixel wide)
• Mask responses
with minus and two with plus signs. * Extreme values
Separating these two blocks is a mask
element with zeros which reduces sensi- Figure 5.6. Edge localization by shifting
tivity to slightly deviating edge direction a ternary correlation mask