Ball localization for RoboCup soccer using convolutional neural networks

Abstract

In RoboCup soccer, ball localization is an important and challenging task, especially since the last change of the rule which allows 50% of the ball’s surface to be of any color or pattern while the rest must remain white. Multi-color balls have changing color histograms and patterns in dependence of the current orientation and movement. This paper presents a neural approach using a convolutional neural network (CNN) to localize the ball in various scenes. CNNs were used in several image recognition tasks, particularly because of their capability to learn invariances in images. In this work we use CNNs to locate a ball by training two output layers, representing the x- and y-coordinates, with normal distributions fitted around the ball. Therefore the network not only locates the ball’s position but also provides an estimation of the noise. The architecture processes the whole image in full size, no sliding-window approach is used.