Adaptive colour classification for RoboCup with Gaussian mixture model

Colour has been used in many computer vision applications, such as image segmentation, object tracking and recognition. The appearance of an image is affected by illumination and so colour-based vision applications have often faced the problem of colours being sensitive to illumination variation. A static colour model can not handle illumination variation and so an adaptive colour model was introduced to deal with dynamic illumination. Our work is motivated by the need for colour classification in robocup research. We have developed an adaptive colour classification algorithm that uses a two component Gaussian Mixture Model (GMM) to model a colour distribution in YUV colour space. The components of this model represent the diffuse and the specular parts of the dichromatic reflectance model. The GMM is derived from classified colour pixels using the standard Expectation-Maximization (EM) algorithm, and the colour model is repeatedly updated with the derived GMM. We propose the novel idea that a GMM with two Gaussian components is an accurate and complete representation of the colour distribution of a dichromatic surface. This work is of practical significance because our adaptive system provides accurate colour classification under variant lighting conditions and it outperforms the previous colour vision system without adversely affecting efficiency.