Long-Term Target Tracking Algorithm Based On Hierarchical Convolution Features

As a mainstream direction in the field of computer vision,target tracking has important research significance and wide commercial value,which has attracted more and more attention from computer vision researchers.In recent years,with the efforts of researchers at home and abroad,high-level target tracking algorithms represented by correlation filtering and deep learning methods have been continuously proposed,but when the target encounters complex conditions such as light changes,scale changes,occlusion,and fast motion,etc.Tracking targets quickly and accurately still poses great challenges.Because convolutional neural networks have excellent feature expression capabilities,convolutional features are used instead of hierarchicalconvolutional features in correlation filtering(HCF),and good experimental results have been achieved.Based on the hierarchical convolution feature correlation filter,this thesis improves on feature extraction,tracking model and fusion processing of multiple trackers.The main contents of this article are as follows:(1)Aiming at the problem of redundancy in the pre-trained convolutional features in the tracking task,a target perception model is designed.Through the first frame of target information,the deep features with target perception are learned,reducing the number of features while alleviating the model overfitting problem.(2)Aiming at HCF effect of the original algorithm in complex environments such as rapid motion,rapid deformation and target occlusion,a feature-based tracking framework is proposed.The feature map is implicitly interpolated from the frequency domain to the continuous domain,and each features branch independently learns a convolution operator composed of a series of spatial regularization continuous filters added in the continuous spatial domain.An adaptive decision fusion strategy is designed to adaptively fuse the target positions of different branches to locate the target.(3)Due to the lack of a re-detection module,HCF cannot track the target again after the target tracking fails.Aiming at the problem that the original algorithm can not adapt to the target and it will appear a long-term tracking problem that completely leaves the fieldof view,a long-term tracking strategy based on the layered convolution feature is proposed.On the basis of the original tracking algorithm,a verification mechanism composed of twin networks is added to correct the tracker in time to make the tracker keep up with the target again,and use the reliability of the tracking result as feedback information to guide the template update of the filter.