Multidimensional Motion Segmentation and Identification

Accurate tracking can facilitate the automatic extraction of metric information from video analysis. Many tracking systems rely on a sufficiently accurate dynamic model. These dynamic models must be either known a priori or learnt. This paper addresses the problem of determining dynamical system models from observed visual motion where it is assumed that the motion cannot be modeled by a single dynamical system. The changes in motion (from one system to another) need to be detected. Previous work has dealt with maintaining multiple hypotheses. For repetitive motion, rather than maintain multiple hypotheses one can learn the dynamic models that apply and identify the changes between the models. Specifically, a method for high dimensional motion segmentation is presented. By using a two-step recursive least square algorithm, break points of system dynamics, at which a model switching must be performed, are predicted. After segmentation, system identification techniques can be used to fit dynamic models.