Regularized approximate residual weighted subsampling for visual tracking

In discriminative tracking algorithms, the accuracy of classifier which relies heavily on the selection of training samples can directly influence the performance of visual tracking. Motivated by above, a tracking algorithm is presented based on regularized approximate residual weighted subsampling in the paper. Through the subsampling procedure, the corrupted samples which exert adverse impacts on the estimated classifier are ensured to be selected infrequently, thus making the classifier trained with the selected sample subset more robust to the noise caused by object appearance variations. Furthermore, an effective model updating strategy is adopted to enhance the flexibility of the tracker to the changes. Compared with some state-of-the-art trackers, our tracking algorithm performs better on a typical benchmark.