Robust Visual Tracking Based on Complementary Diverse Information

Correlation Filters have become the dominant tracking approaches with the state-of-art performance in challenging videos. However, many of them concentrate on stronger feature descriptors or more sophisticated machine learning techniques but ignore the diverse information existing in sequences such as global context, motion states and response maps, consequently leave out lots of valuable clues to tracking. To tackle with this problem and take full advantage of the videos, a succinct tracker is proposed by simply merging response maps inferred by these diverse information. Additionally, to avoid model pollution caused by occluded samples, the fluctuation of response maps are exploited to determine whether to update the model. The experiment results reveal that the combination of the above diverse information with two simple standard features can significantly improve the performance with a gain of 4.2% in mean success rate on ()TB-2015. Our tracker outperforms some recent trackers based on deep features or deep learning frameworks trained with large data set. It demonstrate that diversity and complementarity of tracking information play a crucial part in tracking process.