Localised mixture models in region-based tracking

Abstract: An important problem in many computer vision tasks is the separation of an object from its background. One common strategy is to estimate appearance models of the object and background region. However, if the appearance is spatially varying, simple homogeneous models are often inaccurate. Gaussian mixture models can take multimodal distributions into account, yet they still neglect the positional information. In this paper, we propose localised mixture models (LMMs) and evaluate this idea in the scope of model-based tracking by automatically partitioning the fore- and background into several subregions. In contrast to background subtraction methods, this approach also allows for moving backgrounds. Experiments with a rigid object and the HumanEva-II benchmark show that tracking is remarkably stabilised by the new model.