In soccer, the
ball is the one object on the field that a player must attend to at all
times. If a player looses track of the ball position, it cannot
continue play but must search for it. For this reason, the ball in
RoboCupSoccer is colored in orange, which makes it possible to localize
it in camera images by classifying individual pixels to color classes.
This simple approach has several problems. For example, other orange
objects next to the field might be confused with the ball. Another
problem is motion blur, in particular for humanoid robots capturing
images while walking fast. The presence of orange and green at the same
pixel during exposure leads to a mixture of the two colors: brown, a
common color in the images. Ball detection based solely on color must
fail in such a case.
To overcome these problems, we propose a two-stage system for ball
detection and tracking. First, an extended color class is used to find
ball candidates. Both, color and luminance from small windows around
the candidate locations are classified by a neural network, which has
been trained on a large set of balls and distractors. Thus, in addition
to color, the network can analyze the shape of the object of interest
as well as its shading, including typical highlights and shadows. A
detected ball is tracked in a small window in order to achieve a high
frame rate on a Pocket PC. This also focuses the attention of the
system onto the tracked ball. We evaluated the proposed approach on our
NimbRo KidSize 2006 and 2007 robots. The experiments indicate that the
ball can be reliably detected and confusion with orange non-ball
objects can be avoided.