Guided importance sampling based particle filtering for visual tracking

Linear estimation based sequential importance sampling methods for particle filters are proposed that can be used to model the rapid change of object motion in a video sequence. First a linear least-squares estimation is used to build a proposal function from observations, and then it is extended to a robust linear estimation. These sampling methods give a framework for tracking objects whose motion cannot be well modeled by a prior model. Finally a switching algorithm between the proposed method and the prior model based sampling method is proposed to achieve a filtering of both smooth and rapid evolution of the state. The ability of the proposed method is illustrated on a real video sequence involving a rapidly moving object.