Page 191 - Introduction to Autonomous Mobile Robots

176

Figure 4.49 Chapter 4
Two typical images acquired by the OmniCam catadioptric camera system.

Whole-image features are not designed to identify specific spatial structures such as obsta-
cles or the position of specific landmarks. Rather, they serve as compact representations of
the entire local region. From the perspective of robot localization, the goal is to extract one
or more features from the image that are correlated well with the robot’s position. In other
words, small changes in robot position should cause only small changes to whole-image
features, while large changes in robot position should cause correspondingly large changes
to whole-image features.
We present two techniques for whole-image feature extraction below. The first tech-
nique is another popular application of the image histogramming approach. The resulting
image histogram comprises a set of whole-image features derived directly from the pixel
information of an image. The second technique, tiered extraction, covers approaches in
which a whole-image feature is built by first extracting spatially localized features, then
composing these features together to form a single metafeature.

Direct extraction: image histograms. Recall that we wish to design whole-image fea-
tures that are insensitive to a small amount of robot motion while registering significant
changes for large-scale robot motion. A logical first step in designing a vision-based sensor
for this purpose is to maximize the field of view of the camera. As the field of view
increases, a small-scale structure in the robot’s environment occupies a smaller proportion
of the image, thereby mitigating the impact of individual scene objects on image character-
istics. The catadioptric camera system, now very popular in mobile robotics, offers an
extremely wide field of view [114]. This imaging system consists of a high-quality CCD
camera mounted, together with customized optics, toward a parabolic mirror. The image
provides a 360-degree view of the robot’s environment, as shown in figure 4.49.