Visual Tracking by Assembling Multiple Correlation Filters

In this paper, we present a robust object tracking method by fusing multiple correlation filters which leads to a weighted sum of these classifier vectors. Different from other learning methods which utilize a sparse sampling mechanism to generate training samples, our method adopts a dense sampling strategy for both training and testing which is more effective yet efficient due to the highly structured kernel matrix. A correlation filter pool is established based on the correlation filters trained by historical frames as tracking goes on. We consider the weighted sum of these correlation filters as the final classifier to locate the position of object. We introduce a coefficients optimization scheme by balancing the test errors for all correlation filters and emphasizing the recent frames. Also, a budget mechanism by removing the one which will result in the smallest change to final correlation filter is illustrated to prevent the unlimited increase of filter number. The experiments compare our method with other three state-of-the-art algorithms, demonstrating a robust and encouraging performance of the proposed algorithm.