Data driven visual tracking via representation learning and online multi-class LPBoost learning

摘要

Visual object tracking is a challenging task due to two intractable problems: visual appearance representation and online update model. Existing approaches often operate appearance model based on hand-crafted features with discriminative feature selection. The tracking learning model is formulated as a binary classification. However, some issues remain to be addressed. First, there does not exist sufficient information for online feature selection. Second, these algorithms do not make use of structure information between object and background. In this study, the authors propose an algorithm named data driven tracker with an appearance model which exploits prior visual target representation by binary PCANet. The authors' speed up strategy by binary operation on the convolution filters is efficient for tracking task with little performance loss. They formulate the learning model as multi-class task via online LPBoost. Their data-driven tracking (DDT) algorithm performs favourably on various challenging sequences by evaluating against state-of-the-art trackers.