Combining discriminant embedding and transfer learning for visual tracking

Abstract

Visual tracking can be viewed as a discrimination problem to distinguish the target object from the background. However, it is difficult to get efficient training samples and there is usually a strong similarity between the foreground and the background in reality, which makes it challenging to discriminate the target object from the background. In this paper, we present a tracking method based on the combination of discriminant embedding and transfer learning to tackle the challenge. For one thing, we use the graph embedding method to characterize the relationship between the foreground samples and the background samples, for another we exploit the knowledge of the tracking results in the previous frame to track the next frame. We then learn a subspace by jointly optimizing discriminant embedding and transfer learning into a unified framework. The classifier is constructed on the learned subspace to discriminate the target from the background. Experiments on several video benchmarks demonstrate the effectiveness of our approach.