Study On Visual Tracking Based On Correlation Filters

Visual tracking is a research hotspot in computer vision,and is the basis for video analysis and behavior understanding.In the literature,a plethora of tracking theories,trackers,and datasets have been proposed in recent years.Among them,Correlation Filter(CF)-based tracking has attracted much attention due to its high computational efficiency.Nonetheless,the CF framework still presents several challenges,such as boundary effects and filter degradation.In view of the above challenges,this thesis combines deep neural networks to carry out in-depth research on CF tracking,analyzes the shortcomings of existing models,and proposes developed trackers.The main contributions and innovations of this thesis are summarized as follows:1.To handle the boundary effect caused by the assumption of sample periodicity and the filter degradation caused by the linear update of the filter,an adaptive spatio-temporal CF-based tracking algorithm is proposed in this thesis.A spatio-temporal regularization based on the change of the target appearance is learned,which adaptively adjusts the spatial weight and temporal coefficient to effectively deal with boundary effects and filter degradation.In addition,the objective function can be quickly optimized using alternating direction method of multipliers.2.To handle the sudden change of tracking response and aberrant distribution of the filter caused by aberrance,a spatio-temporal joint aberrance suppressed CF-based tracking algorithm is proposed in this thesis.A dynamic spatial regularization is learned,which imposes different penalties on the filter coefficients at each pixel in the training region to obtain more reliable filter coefficients and alleviate boundary effects.A dynamic temporal regularization is learned,which balances the differences between current and previous filters to maintain the manifold distribution between filters.An aberrance suppression strategy is introduced,which penalizes the response difference between adjacent frames to limit sudden change of response and improve the accuracy of target localization.3.To handle the poor real-time performance and model drift of the deep-based CF tracking,an accelerated dual-aware CF-based tracking algorithm is proposed in this thesis.A dual model structure of translation-scale filter is designed to avoid the time-consuming problem of multiscale deep feature extraction and improve the accuracy of scale estimation.A sparsedecision model update strategy is proposed,which reduces the computational complexity of the algorithm,and dynamically adjusts the learning rate of the training sample set to filter out aberrant samples and improve the discrimination of the model.