Local background-aware target tracking

Classical visual tracking methods usually only adopt target information to describe the target. In practice, local background information around the target also influences the tracking process. In this paper, we first introduce the local background information into target description, and represent it as a set of weighted feature points. After that, the posterior probability of each point in the search-region is calculated by incorporating the target observation obtained by K nearest neighbor (KNN) algorithm and the Gaussian prior distribution of the target. Finally, the mean shift algorithm is used to estimate the target state. The proposed method has the following two advantages: 1) The local background information is integrated into the target description, which enhances the target model. Thereby, the discriminative ability is promoted in the tracking process, which further makes the tracker more robust and accurate. 2) In the initialization stage, the mean-shift is applied to relocating the target, which can solve the problem that the tracking algorithm is prone to failure in inexact initialization. Extensive experiments in different video sequences are conducted to evaluate our approach qualitatively and quantitatively. The results show that our method holds high tracking accuracy and stability, especially when the target is roughly initialized.