Using the visual attention system VOCUS combined with a fast classifier, we have designed a robust ball detection system with a very low misclassification rate, even in complex, cluttered images. Due to the use of an edge detection Sobel filter and a threshold to preprocess the training images for the cascade, the classifier is color-invariant, leaving the color to be learned by the attention system. Assuming short-term prior knowledge about the ball to be used for a RoboCup match, VOCUS is quickly adjusted to the ball with very few images.
The success of the algorithm is reached by only searching for balls in regions hypothesized by the attention algorithm to contain the ball, thereby eliminating false positives. Although the algorithm misses a few balls, what we are concerned with is how it will perform in the RoboCup environment. In this case, the reliability of the algorithm seems to be sufficient. Even if the ball is not detected in one in every 5 pictures, for example, the robot will still be able to follow it quite confidently.
Needless to say, much work remains to be done: As the detection of regions of interest is currently relatively slow compared to the ball detection, the next step is to work on increasing the efficiency of the attention system and therefore of the whole detection scheme. In addition it is planned to enhance the presented algorithms by adding time dependent behavior either by using standard tracking with particle filters or by using a time dependent attention control.