Online similarity learning for visual tracking

Abstract

Incorporating metric learning in visual tracking applications has been demonstrated to be able to improve tracking performance. However, the optimal metric is mainly derived based on annotated feature vectors by studying their magnitude and intersection angle. In complex scenarios, the magnitude of feature samples may change drastically, confining the matching performance of the distance metric. Moreover, most distance learning methods are optimized in a time-consuming iterative way, which limits their applications in real-time visual tracking. To address these problems, this paper proposes a novel metric called sphere similarity metric, which normalizes the magnitude of feature vectors and measures the distance between pairs of vectors by their intersection angle. Such metric is robust even when the magnitude of feature vectors changes drastically in complex scenarios. We formulate the proposed metric by a convex matrix function, which does not require adapting samples’ magnitude iteratively, and can be optimized in a closed form with low computational complexity. Additionally, the proposed similarity metric can be learned in an online manner, which accelerates the learning process and improves the tracking accuracy in visual tracking applications. Experimental results on synthetic data and benchmark video sequences show that the proposed metric learning method achieves better classification accuracy, and its application in visual tracking outperforms the state-of-the-art tracking methods.