Research On Structured Target Tracking Algorithm Based On Sparse Representation And Kernel Method

As a fundamental research problem in computer vision,visual object tracking has been received a lot of attention from academia and industry.Researchers have proposed many excellent tracking algorithms.However,visual object tracking is still a challenging research topic because of the complexity of tracking scenes.In this work,we studied sparse representation(SR)tracking algorithms and kernel correlation filter(KCF)algorithms.We point out the deficiencies in both algorithms,and this work proposes improvements and verify them through experiments.The innovations of this work are listed as follows:(1)Under the sparse representation visual tracking framework,an adaptive weighted structured tracking algorithm is proposed in this thesis.The traditional sparse representation algorithm model the whole object which is not conducive to the tracking of the target object under the occlusion situation.This work adopts the structured local sparse appearance model,and proposes the local patch weighting strategy and the patch discarding strategy,which can deal with the occlusion situation well.In addition,this thesis proposes an adaptive dictionary template update method,and determine whether it is necessary to update the template set according to the reconstruction error,so as to adapt to the appearance change of the target in time and prevent the template set from being polluted by the error templates.Experimental results verify that the tracking performance of the algorithm has significantly improved compared with the basic sparse representation algorithm while dealing with target occlusion on the object tracking benchmark(OTB)test dataset.(2)Under the kernel correlation filter framework,an adaptive spatial context tracking algorithm is proposed in this thesis.Firstly,for the deficiencies of KCF training filter by very limited background information,we add spatial context information in the training process to enhance the discriminative ability of the filter.Second,for the scale change of the target,a scale filter is trained to accurately estimate the target scale.Finally,the traditional KCF only uses hand-craft features with weak representation capabilities.This work uses convolution features from different layers,and combines the advantages of convolution features from deep and shallow layers to estimate the target position accurately.Experimental results verify that the tracking performance of the proposed algorithm has significantly improved compared with the basic KCF tracker while dealing with various complex scenarios on the OTB test dataset.(3)This thesis combines sparse representation with kernel correlation filter,and proposes a structured local sparse tracking algorithm.The KCF algorithm uses the virtual cyclic samples training filter during the training phase,causing certain errors in the filter.In the detection phase,the virtual cyclic sample and the filter with errors are used together to detect the target object,resulting in inaccurate tracking results.In this work,we make improvements during detection phase and combine KCF with SR.In detection phase,the samples with large response value are selected as the real candidate samples,and we use these real samples to calculate response values.Besides,the structured local sparse appearance model is used to model these real samples and calculate the observed likelihood values.An adaptive weighting method is designed to weight the response values and the observed likelihood values,and sum them.The candidate sample position with the largest summation result is the position of the tracking object,then the scale filter is used to estimate the target scale to obtain the final tracking result.This algorithm can reduce the errors caused by the use of the cyclic samples to a certain extent.Experimental results verify that the proposed algorithm has excellent tracking performance compared with the KCF trackers and deep learning trackers while dealing with various complex scenarios on the OTB test dataset.