Skip to Main Content
In this paper, we present a novel on-line probabilistic generative model that simultaneously deals with both the clustering and the tracking of an unknown number of moving objects. The proposed model assumes that i) time series data are composed of a time-varying number of objects and that ii) each object is governed by a mixture of an unknown number of different patterns of dynamics. The problem of learning patterns of dynamics is formulated as the clustering of tracked objects based on a nonparametric Bayesian model with conjugate priors, and this clustering in turn improves the tracking. We present a particle filter for posterior estimation of simultaneous clustering and tracking. Through experiments with synthetic and real movie data, we confirmed that the proposed model successfully learned the hidden cluster patterns and obtained better tracking results than conventional models without clustering.
Date of Conference: 23-28 June 2008