Scale adaptive correlation tracking based on convolutional features

In recent years, several correlation tracking algorithms have been proposed exploiting hierarchical features from deep convolutional neural networks. However, most of these methods focus on utilizing the hierarchical features for target translation and use a fixed-size searching window throughout a sequence. Because of neglecting the changes of target scale, these algorithms may import error to the model and lead to drifting. Moreover, numerous factors like fast motion and heavy occlusion can induce instability of translation model which may result in the tracking failure. In this paper, we propose a novel scale adaptive tracking algorithm based on hierarchical CNN features, which learns correlation filters to locate the target and constructs a target pyramid around the estimated target position for scale estimation. In case of tracking failure, we generate an online detector of random fern classifier and activate it to re-detect the target. To evaluate the tracking algorithm, extensive experiments are conducted on a benchmark with 100 video sequences. The tracking result demonstrate that the hierarchical CNN features are well fit to handle sequences with scale variation, motion blur and illumination variation. And the online re-detection is of great importance in tracking failures caused by fast motion and heavy occlusion. The evaluation results show that our tracker outperforms the state-of-the-art methods by a huge margin.