Unsupervised Skin Colour Modelling for Hand Segmentation

Hand detection is a fundamental step in many practical applications as gesture recognition, video surveillance, multimodal machine interface and so on. In this paper a colour based hand segmentation methodology able to work without any a priori knowledge about colour, lighting and spatial conditions is proposed. The methodology consists of four steps: motion analysis, face detection, colour skin modelling and hand segmentation. At first motion analysis is performed by adaptive background subtraction in the RGB colour space. The moving pixels are then supplied as, input to an unsupervised clustering algorithm that uses both colour and spatial position information. Topological and geometrical constraints are used to detect the face cluster which is used to build a colour skin model by multivariate Gaussian functions in the normalized RGB colour space. This. statistical model is finally used to retrieve the hands position both in the current and in the forthcoming frames.