Robust Object Tracking via Graph-based Transductive Learning with Subspace Representation

Although many tracking algorithms have been studied in recent years, target tracking is still a very basic research topic. In this study, we propose a novel robust object tracking via graph-based transductive learning with subspace representation (GTLSR). Firstly, probabilistic hypergraph ranking theory is developed to capture the local affinity information amount all vertices. Then, we transform the object tracking into the transductive learning problem based on Bayesian inference framework. Third, to deal with the occlusions of target, we use subspace representation to constrain the indication vector of template set. Finally, a dynamic updating strategy based on double threshold is proposed to solve the challenges of pose change and occlusion. The extensive experiments reflect that the proposed GTLSR outperforms other baseline methods in accuracy and robustness.