Object tracking with double-dictionary appearance model

Abstract. Dictionary learning has previously been applied to target tracking across images in video sequences. However, most trackers that use dictionary learning neglect to make optimal use of the representation coefficients to locate the target. This increases the possibility of losing the target in the presence of similar objects, or in case occlusion or rotation occurs. We propose an effective object-tracking method based on a double-dictionary appearance model under a particle filter framework. We employ a double dictionary by training template features to represent the target. This representation not only exploits the relationship between the candidate and target but also represents the target more accurately with minimal residual. We also introduce a simple and effective strategy to update the template to reduce the influence of occlusion, rotation, and drift. Experiments on challenging sequences showed that the proposed algorithm performs favorably against the state-of-the-art methods in terms of several comparative metrics.