Visual tracking via adaptive multi-task feature learning with calibration and identification

Recently multi-task feature learning has become a widely applied approach for visual tracking, since it is benefited from the shared features across tasks. However, selecting features appropriately from multiple tasks is still a challenging problem due to the complex variation of the appearance of moving objects, which influences not only the features of single task but also the relationships between the features of multiple tasks. To address this problem, this paper presents a novel sparse learning model for selecting multi-task features adaptively. Compared to the existing multi-task models, the proposed model is capable of both calibrating the loss function according to the noise level of a task to keep its specific features, and identifying the relevant and irrelevant (outlier) tasks simultaneously by decomposing the regularized matrix into two specified structures. The proposed model allows to preserve specific features of individual tasks via calibration and to exploit sparse pattern over the relevant task via identification. Empirical evaluations demonstrate that the proposed method has better performance than a number of the state-of-the-art trackers on available public image sequences.