
92
4.
Vision
trated. Basically, the robot is looking for "staple" shaped
boundaries (see the last panel of Figure 4-11). The robot
considers an object to be a can if it has a flat top, and two fairly
straight sides of equal length.
The first step of the actual visual processing is to convert the
disparity map into a binary image. Any pixel for which a true
depth reading obtained is marked as a one, all the others are
zeroed. This simplification is possible because we do not use
any depth information in the recognition phase of can finding.
Next, to combat streaking, we eliminate any features narrower
than 6 pixels. This should not remove an ...