Research on deep correlation filter tracking based on channel importance

Abstract

Abstract Correlation filter tracking requires little prior knowledge of the tracking target (e.g., the shape, and the posture) but has a fast-tracking speed. The deep features extracted by the deep convolutional neural network have strong representation ability, so the tracking method based on the combination of correlation filter and deep convolutional neural network, named as deep correlation filter tracking, is a hot issue in the field of target tracking at present. However, the deep convolutional neural network largely restricts the real-time performance of the deep correlation filter tracking because of its complex network structure and heavy computation burden. To balance the contradiction between tracking speed and tracking accuracy, a new channel importance is defined and the channel importance based method of how to select the important channels is given in this paper. And then, a deep correlation filter tracking method based on channel importance is proposed to lighten the feature network, reduce the computation load and improve the tracking speed under the premise of ensuring the tracking accuracy. In the process of tracking, the structural similarity index measurement (SSIM) of the predicted tracking target in two consecutive frames is calculated in real-time. Based on the SSIM, determine whether the feature network needs to be updated, and decide whether the tracking fails. If the feature network needs to be updated, the feature network will be updated online while the tracking is on. If the tracking fails, the target will be searched again, and the tracking is recovered from the failure. The tracking algorithm proposed in this paper is tested on the OTB2013 data set, and the experiment shows that the tracking algorithm designed in this paper can improve the real-time performance while meeting the requirement of tracking accuracy. The online update of the feature network can make the network adapt to the complex background and target changes to improve tracking accuracy; In the case of tracking failure, the re-tracking module can search for the target again and resume tracking given that the target is always present.

Keywords

• Object tracking
• Deep correlation filter
• Channel importance
• SSIM