Spatial Graph Regularized Correlation Filters for Visual Object Tracking

In Visual object tracking (VOT), Discriminative correlation filters trackers have achieved promising results for VOT in many complex scenarios. However, because of the unwanted boundary effects and lack of structural constraints, these methods suffer from performance degradation. In the current work, we propose a spatial graph-regularized correlation filter for robust VOT. In this method, we transform the circulant shifted target samples to a particular subspace such that the target and the background become linearly separable. For this purpose, we encode pairwise similarities among the circulant shifted target samples as a spatial graph via a learnt correlation filter constrained to act as an eigenvector of the Laplacian of this spatial graph. We propose an objective function which incorporates this spatial constraint into the DCFs learning framework, which we solve using ADMM with a closed-form solution. Evaluated on a set of four datasets, our framework showed superior performance when compared to competitive state-of-the-art (SOTA) methods.