RoboCup Small-Size League: Using Neural Networks to Learn Color Segmentation during Visual Processing

Optimizing vision processing is crucial for real-time performance of robots in RoboCup's Small-Size League (SSL). We describe in this paper our current approach to improve visual processing in ITAM's Eagle Knights SSL team. We describe our use of a neural network to classify camera image pixels to a discrete set of color classes that is robust under different light conditions. We show how we can improve the recall time of the neural network to achieve vision processing of over 30 fps using high resolution images. We present our solution and compare to previous methods showing improvements in real time image segmentation and varying light conditions.