Toward automatic colour calibration using machine learning

Abstract

Robot vision can be seriously affected by variations in lighting conditions in the environment. Therefore, vision systems need to be recalibrated whenever lighting conditions change, otherwise they fail to recognize objects or classify them incorrectly. This paper describes experiments aimed at overcoming the problem of vision recalibration. We use Sony ERS210A robots, which are used in RoboCup, the international robot soccer competition. These robots are equipped with colour cameras. The robots are trained to recognize colours under different lightning conditions using machine learning. Learned knowledge is then used for object recognition. In addition, the robots are trained to recognize objects using colour and other object attributes. Incorporating leaning into the vision system facilitates object recognition and provides the ability to automatically adapt vision to new environments.