Video Tracking via Tensor Neighborhood Preserving Discriminant Embedding

In a real surveillance scenario, tracking an object usually interfered by the background information. To deal with this problem, this paper proposed a video tracking algorithm based on tensor neighborhood preserving discriminant embedding. The neighborhood relationships of an object within object class and background class are reasonable described by the object image patches similarities which are defined by histograms of oriented gradients. In order to distinguish between the object and background, we formulate an discriminant objective function that maximizing the scatters of object within object class while minimizing the scatters of object with background class, meanwhile maintaining the same neighborhood topological structure in lower dimensional tensor subspace. Finally, we can get the optimal estimate of the object state through Bayesian estimation framework. Experimental evaluations against two state-of-the-art tracking methods demonstrate the robustness and effectiveness of the proposed algorithm.