Sequential Importance Sampling for Visual Tracking Reconsidered

We consider the task of filtering dynamical systems observed in noise by means of sequential importance sampling when the proposal is restricted to the innovation components of the state. It is argued that the unmodified sequential importance sampling/resampling (SIR) algorithm may yield high variance estimates of the posterior in this case, resulting in poor performance when e.g. in visual tracking one tries to build a SIR algorithm on the top of the output of a color blob detector. A new method that associates the innovations sampled from the proposal and the particles in a separate computational step is proposed. The method is shown to outperform the unmodified SIR algorithm in a series of vision based object tracking experiments, both in terms of accuracy and robustness.