Robust Single-object Visual Tracking Framework via Fully Convolutional Siamese Network with Correlation Filter

Recently, there are two kinds of methods in the research field of visual tracking, one is correlation filtering, the other is siamese network. In order to fully utilize the advantages of these two framework, this paper proposed a robust single-object visual tracking framework based on fully convolutional siamese network and correlation filter. Two branches are constructed in this network, namely training template branch and searching template branch, which are share the convolutional layer parameters. Each branch is divided into two sub-branches. In the sub-branches of the training template, two correlation filtering layers are embedded to replace the traditional convolutional layer. Finally, the location of object to be tracked is calculated by constructing response graph. Extensive experiments on challenging benchmark show that the tracking framework in this paper can further improve the accuracy.