Proposed is a novel clustering technique based on kernel methods. The geometric properties of normalised kernel spaces are exploited to automatically detect the correct number of clusters, thus avoiding the requirement of an initial estimate of this parameter, as required instead in many popular algorithms.