Temporal–spatial consistency of self-adaptive target response for long-term correlation filter tracking

Abstract

Owning to the impressive performance and high speed of the correlation filter (CF) visual tracking technology, the CF-based tracking gets a significant amount of attentions and obtains a large number of researches. But the traditional CF-based trackers have two inherent issues needed to be solved. On the one hand, the boundary effect caused by the circulant structure will affect the performance. On the other hand, the target response is assumed as Gaussian response and the value is fixed. In order to deal with the boundary effect, a spatially regularized method by imposing space penalties to the spatial coefficients is proposed. About the fixed target response, some researchers propose a framework that can adaptively change the response to deal with it, but the two issues only be coped with separately. We propose a tracker that utilizes the two merits jointly to win a better performance and a more robustness model which can cope with the complex environment to achieve a long-term tracking. We use an iteration solution to solve the joint cost function and ADMM algorithm to solve temporal–spatial consistency problem which can obtain a global solution and fast computation. We evaluate our method on the well-known visual tracking benchmark dataset called OTB50 and get a competitive performance result compared with the relevant trackers.