Spatial-temporal adaptive feature weighted correlation filter for visual tracking

In recent years, the discriminative correlation filter (DCF) tracking method has attracted considerable attention for its promising results and high efficiency. However, it introduces unexpected background information and is susceptible to distractors, owing to the larger training patches it uses. This paper proposes a correlation filter-based visual tracking approach that integrates spatial-temporal adaptive feature weights into the DCF method to derive an efficient solution. The feature weights, which are adaptively adjusted according to the target likelihood and prior knowledge, effectively enhance the target and suppress the background. The correlation filter in our formulation is learned based on weighted features. By assigning small weights to the background and large weights to the target, the tracker is less likely to drift. To preserve the circulant property in the DCF formulation, a new variable called the pseudo correlation filter is introduced to reformulate and solve the cost function efficiently by employing the iterative Gauss-Seidel method. The proposed method is tested on both handcrafted and deep features. Extensive experiments on three benchmark datasets demonstrate that the proposed tracker achieves favorable performance compared with recent state-of-the-art methods.