Visual vehicle tracking based on an appearance generative model

This paper consider the problem of handling appearance variability in visual tracking and proposes an appearance generative model for visual vehicle tracking. The generative model is used to adaptively generate and update the appearance templates during visual tracking. The appearance templates are efficiently represented in a low dimensional eigen subspace learned from pre-acquired templates and are parameterized by two pose parameters of a target object. The adaptive template updating is made by particle filtering in which the particles represents the appearance templates. In experiments with real image sequences, we show the effectiveness of the proposed method.