Ensemble Tracking Based on Randomized Trees

Object tracking is an active yet challenging research topic in computer vision. Recently, a trend to treat the problem as a classification problem is boom. By such a paradigm, a discriminative classifier is trained and updated during tracking procedure. In this paper, the ensemble of randomized trees such as random forests or extremely randomized trees is employed to construct a discriminative appearance model to accomplish tracking task. Benefited from the noise insensitivity and operation efficiency of randomized trees, the appearance model used for tracking can be efficiently updated through growing new trees to substitute the degraded ones. Meanwhile, mean shift is introduced to locate the object in each newly arrived frame. Extensive experiments are performed to compare the proposed algorithm with four wellknown tracking algorithms on several challenging video sequences. Convincing results demonstrate that the proposed tracker manages to handle illumination changes and pose variations.