Revision history [back]

I am not sure if haar cascades are a good idea here. First of all, the robot would have to be given a lot of examples before the haar cascade would be useful. Secondly, training a haar cascade requires a lot of time on a desktop machine, so that would take ages on the Nao. Thirdly, you would have to have as many cascades as there are characters. Thus, the recognition phase would be quite long, as each cascade would have to be used on each unknown character. I don't know if I'm right, but I would use a more classical approach: adaptive thresholdig, some feature extraction, a classifier training (something quick, most likely random trees). If the selected features are good, the classifier should be better and better with each new example.