Conclusions

The Gentle Ada Boost algorithm uses Classification and Regression trees (CARTs) with four splits to construct a cascade of classifiers to detect Haar-like features in integral images. To ensure the color-invariance of the input images, they are first preprocessed by applying an edge detection Sobel filter to each of the images and passing them through a threshold to rid them of all their color information. This has proven to be a relatively successful technique to be used by the autonomous mobile robot Kurt3D to detect footballs.

It has been found that there are quite a few parameters that need to be adjusted to get satisfactory results from the algorithms, such as the filtering algorithm used, its parameters, the number of splits in the CART, the number of training images used, and above all, the selection of training images. The aim of this paper was not to discuss the details of all these parameters, but the most important ones have been mentioned, and their influence on the results shown.

Although the results may not seem very positive, 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.

There still remains a lot of room for improvement, though, especially concerning the false detection of all round objects (see Fig. 9), as well as undetected footballs with lots of background noise, or partially visible footballs. This could be achieved by combining our approach with other techniques or by integrating 3D images to include depth information. Another idea is to use attention algorithms - assuming the color of the football is previously known - to define regions of interest in order in which to search for balls in order to eliminate false positives. It is planned to work on this in the near future.