Mode-based multi-hypothesis head tracking using parametric contours

The paper describes a probabilistic mode-based multi-hypothesis tracking (MHT) algorithm. The modes are the local maximums refined from initial samples in a parametric state space. Because the modes are highly representative, this technique allows us to use a small number of hypotheses to effectively model nonlinear probabilistic distributions. To ensure real-time tracking performance, we propose a novel parametric causal contour model and an efficient dynamic programming scheme to refine the initial contours to nearby modes. Furthermore, to overcome the common drawback of conventional MHT techniques, i.e., producing only the maximum likelihood estimates instead of the desired posterior, we introduce the highly effective importance sampling framework into MHT, and develop a novel procedure to estimate the posterior from the importance function. Experiments on a challenging real-world video sequence demonstrate that the proposed tracking technique is both robust in complex environment (e.g., clutter background and partial occlusion) and efficient in computation