ℓ1-Graph based semi-supervised learning for robust and efficient object tracking

Online discriminative learning methods have been shown to give promising results in visual tracking. However, the shortage of positive examples representing the object can degrade the classifier. To handle this problem, we propose a ℓ₁-graph based semi-supervised object tracking algorithm, which make full uses of the intrinsic manifold structure of the dataset including both labeled and unlabeled instances to obtain a better classifier. We first extract positive and negative examples as labeled templates from the previous few frames and draw candidates with a particle filter. Then the ℓ₁-graph is constructed based on all templates and candidates. The similarities between the templates and candidates are evaluated over ℓ₁-graph. Lastly, the tracking result is employed to update the ℓ₁-graph. Empirical results on challenging video sequences demonstrate the superior performance of our method in robustness and accuracy to state-of-the-art methods in the literatures.