Abstract:

In order to alleviate the effect of limited labeled instances on the supervised classifier in object tracking, a new method based on the graph based semisupervised learning framework is proposed for object tracking. Firstly, two kinds of positive instances representing the long term and shortterm information about the object respectively and negative instances representing the background are extracted around the tracking results of the previous few frame. Meanwhile, the candidates are sampled by a particle filter as unlabeled instances. Secondly, each labeled or unlabeled instance is divided into several overlapped patches with a spatial layout. The corresponding patches of all these instances are gathered together to construct a graph. The similarity scores of the candidates are evaluated independently over each graph. Finally, these similarities from all graphs are combined to find the tracking result and update the graphs. Empirical results on challenging video sequences demonstrate the superior performance of the proposed method in robustness and accuracy to state of the art methods in the literature.