Online maintaining appearance model using particle filter

Tracking by foreground matching heavily depends on the appearance model to establish object correspondences among frames and essentially, the appearance model should encode both the difference part between the object and background to guarantee the robustness and the stable part to ensure tracking consistency. This paper provides a solution for online maintaining appearance models by adjusting features in the model. Object appearance is co-modeled by a subset of Haar features selected from the over-complete feature dictionary which encodes the discriminative part of object appearance and the color histogram which describes the stable appearance. During the particle filtering process, feature values both from background patches and object observations are sampled efficiently by the aid of "foreground" and "background" particles respectively. Based on these sampled values, top-ranked discriminative features are added and invalid features are removed out to ensure the object being distinguishable from current background according to the evolving appearance model. The tracker based on this online appearance model maintaining technique has been tested on people and car tracking tasks and promising experimental results are obtained.