Real-time object tracking via optimal feature subspace

In this paper, we present a real-time tracking approach based on the Optimal Feature Subspace (OFS). OFS is an optimal subspace of a random feature space, which can best represent the target and making it most distinguished in the whole scene. Initially, we randomly crop patches inside the bounding box to generate an efficient feature template set. Then a greedy algorithm fusing the cues of both target and background is proposed to seek the OFS at every frame. In the forthcoming frame, considering the correlation of different dimensions, we compute the Mahalanobis distance of candidate patches to the appearance model in the obtained subspace to locate the target. The experimental results on several challenging video clips demonstrate that our approach outperforms the state-of-the-art methods, in terms of both speed and robustness.