Page 249 - Introduction to AI Robotics

232
6 Common Sensing Techniques for Reactive Robots

Figure 6.19 A Denning mobile robot using a color histogram to play tag with a
poster of Sylvester and Tweety.

6.7.1 Stereo camera pairs

Using two cameras to extract range data is often referred to as range from
stereo, stereo disparity, binocular vision, or just plain “stereo.” One way to ex-
tract depth is to try to superimpose a camera over each eye as in Fig. 6.20a.
Each camera finds the same point in each image, turns itself to center that
point in the image, then measures the relative angle. The cameras are known
STEREO PAIR as the stereo pair.
This method has two challenges. The first is that it is hard to design and
build a mechanism which can precisely move to verge on the points. (It is
even harder to design and build an inexpensive vergence mechanism.) The
second challenge is even more fundamental: how does the robot know that it
is looking at the same point in both images? This problem is referred to as the
CORRESPONDENCE correspondence problem, since the task is to find a point in one image that cor-
responds to a point in the other image. A common approach is to identify
“interesting” or potentially uniquely valued pixels in the image, such as very
bright or dark spots or edges. The algorithm that selects interesting pixels is
INTEREST OPERATOR called an interest operator. Since even minute changes in lighting make a dif-
ference in the image, there is no guarantee that the two images, even though
acquired at precisely the same time from two cameras, will “see” the same
values for corresponding pixels. Therefore, interest operator algorithms usu-
ally return a list of interesting pixels, not just one, and a matching algorithm
tries to find the best correspondence between all of them. After the interest