Adaptive Color Mapping for NAO Robot Using Neural Network

While playing soccer, the main task of the robot vision system is identifying and tracking objects such as ball, goals, teammate robots and opponent robots. The basis of many object identification methods, particularly those in soccer robots and RoboCup environment, is using algorithms based on pixel color properties. One of the major problems of these robots in RoboCup environment is changes in lighting conditions of match environment and in turn difficulty in identification of environment colors and objects in order to segment the image and identify the objects. In this paper, a pixel color-base identification method has been suggested using a neural network for recognizing the pixels related to each object. The neural network used in this study has 6 output neurons for identifying 6 classes of signs including ball, goal, field, field lines, teammate robot and opponent robot. This proposed method has been tested on over 1000 frames of images received by robot’s camera and different data sets and has revealed an appropriate performance with identification rate of over 90% and error rate of 4%.