Deep tracking using double-correlation filters by membership weighted decision

摘要

Correlation filters are well-known for tracking robustness and accuracy while convolutional neural networks (CNN) is famous for representation learning capability. However, how to combine them to further boost tracking performance remains an open problem. In this paper, we are resolved to derive a more compact double-correlation filter and incorporate an ensemble of double-correlation filters in a membership-based decision fashion where filters are trained on features obtained from different layers of a CNN respectively. The novel double-correlation filter is constructed by maximizing the similarity between the Gaussian-shape label and the correlation of template and training samples, producing a more concise solution that means more computational efficiency. Multiple filters are learned based on multiple-layer CNN features obtained. The final tracking prediction is a membership-weighted decision where membership of each tracker, which is computed according to their performance in previous frames, shows how close a weak tracker-s result is to the truth. Hence our framework not only combines correlation filters and CNN together, but also fully utilizes both deep features providing semantic information to distinguish target from background and their shallow counterparts retaining details beneficial for precise localization. We experiment on benchmark OTB and VOT where our algorithm demonstrates competitive performance versus other state-of-the-art trackers.