Page 151 - Introduction to Autonomous Mobile Robots

136

Figure 4.25 Chapter 4
The SVM module mounted on EPFL’s Shrimp robot.

Many zero crossings do lie at edges in images, but their occurrence is somewhat broader
than that. An interesting characteristic of zero crossings is that they are very sharply
defined, covering just one “pixel” width in the filtered image. The accuracy can even be
further enhanced by using interpolation to establish the position of the zero crossing with
subpixel accuracy. All told, the accuracy of the zero crossing features in ZLoG have made
them the preferred features in state-of-the-art stereo depth recovery algorithms.
Figure 4.24 shows on an example the various steps required to extract depth information
from a stereo image.
Several commercial stereo vision depth recovery sensors have been available for
researchers over the past 10 years. A popular unit in mobile robotics today is the digital
stereo head (or SVM) from Videre Design shown in figure 4.25.
The SVM uses the LoG operator, following it by tessellating the resulting array into sub-
regions within which the sum of absolute values is computed. The correspondence problem
is solved at the level of these subregions, a process called area correlation, and after cor-
respondence is solved the results are interpolated to one-fourth pixel precision. An impor-
tant feature of the SVM is that it produces not just a depth map but distinct measures of