Page 248 - Introduction to AI Robotics

231
6.7 Range from Vision
Because the color histogram of a current image can be matched with an-
other image, the technique appears to be model-based, or recognition. But
reactive systems do not permit recognition types of perception. Is this a con-
tradiction? No; a color histogram is an example of a local, behavior-specific
representation which can be directly extracted from the environment. For
example, a robot could be shown a Barney doll with its distinct purple color
with green belly as the percept for the goal for a move-to-goal behavior.
However, the robot will follow a purple triangle with a green region, be-
cause the ratio of colors is the same. There is no memory and no inference,
just a more complex stimulus.
Note that the intersection can be considered to be a measure of the strength
of the stimulus, which is helpful in reactive robotics. In one set of experi-
ments, a robot was presented a poster of Sylvester and Tweety. It learned the
histogram, then after learning the object (e.g., fixating on it), it would begin to
move towards it, playing a game of tag as a person moved the poster around.
The robot used a simple attractive potential fields-based move-to-goal be-
havior, where the perceptual schema provided the location of the poster and
the percent intersection. The motor schema used the location to compute the
direction to the poster, but the intersection influenced the magnitude of the
output vector. If the person moved the poster into a dark area or turned it at
an angle, the intersection would be low and the robot would move slower.
If the match was strong, the robot would speed up. Overall, it produced a
very dog-like behavior where the robot appeared to play tag quickly (and
happily) until the human made it too difficult. Then if the human moved the
poster back to a more favorable position, the robot would resume playing
with no hard feelings.

6.7 Range from Vision

An important topic in computer vision is how to extract range information.
Humans extract depth from vision. In most cases, though not always, depth
perception is due to having two eyes and being able to triangulate as shown
STEREOPSIS in Fig. 6.20, also known as stereopsis. Other times, perception of depth is
OPTIC FLOW related to optic flow and simple cues such as shadows, texture, and expected
size of objects. This section covers three types of vision sensors which are
commonly used to create an image representing a depth map: stereo camera
pairs, light stripers,and laser ranger finders.