Autonomous color learning in an artificial environment

Abstract

In our application, the RoboCup soccer competition, we are interested whether certain objects (ball, beacons, goals) are present in the field of view. RoboCup soccer is a color-coded environment, where highly saturated colors are used to ease the recognition of those objects essential for RoboCup soccer. So the color of the object, as classified by a human, is the natural feature space to be used as the first step of processing the stream of images. Humans are able to classify the field as green in a color image under a wide variety of lighting conditions, and a robot should be able to learn the same mapping from the 3-dimensional color space to a few color classes. The robots in the 4-Legged League record images in the 3-dimensional YUV color space. A standard approach in color-invariant image processing is to convert such images to a more robust color space (see for an overview [2]). Unfortunately such transformations which are for instance based on Gaussian convolution are out of scope for our robotic platform. We used another machine learning technique. To find the mapping from the 3-dimensional color space to the color classes, an analysis of the feature space is needed. Objects which humans would classify as the same color should be visible as clouds of ’similar’ pixels. It is the task of the algorithm to separate those clouds into clusters. While there is a wide variety clustering techniques available (see for an overview [3]), complex algorithms are out of scope due to the limited computational resources of the RoboCup robots.