Research On Correlation Filter Tracking Algorithm Based On Lightweight CNN

Video target tracking is one of the important research branches of computer vision,which is widely used in science and technology,national defense construction,automatic driving and other important economic fields.Video target tracking algorithm based on correlation filtering has been studied and widely used because of its fast and effective characteristics.In recent years,convolution feature has been widely used in correlation filtering algorithm because of its effective image representation ability.Although new research methods and contents are proposed by researchers,the problem of high number of parameters and high amount of computation of feature extraction network also exists in the target tracking algorithm based on correlation filtering.In order to solve the problem of too much parameter and calculation in feature extraction of target tracking algorithm based on correlation filtering,this paper analyzes and improves the feature extraction of the algorithm based on ECO algorithm,and proposes a new algorithm based on lightweight CNN.The main work and research contents of this paper are as follows:(1)Aiming at the problem of high number of parameters and high amount of computation in feature extraction network,a novel feature extraction network of Ghost Net based on iterative attention feature fusion is proposed.Firstly,we use the Ghost Net as the backbone of feature extraction network.Secondly,the iterative attention feature method integrated to increase the context awareness of features.Finally,we build the feature extraction network,namely IAFF-Ghost Net.The experimental results show that compared with RESNET,the parameters of the feature extraction network are reduced by 77.7% and the calculation is reduced by 91.8%,which reduces the calculation and parameters of the whole feature extraction network.(2)Aiming at the problem of high number of network parameters and high amount of computation in feature extraction of target tracking algorithm based on correlation filtering,an improved IAFF-GECT target tracking algorithm based on ECO algorithm is proposed.Firstly,the IAFF-Ghost Net network is used as the feature extraction network to extract the convolution features of the shallow and deep layers of the images.Secondly,the convolution features and manual features are interpolated and convoluted with the current filter in the Fourier domain to realize the target location.Finally,the conjugate gradient algorithm is used to optimize the loss function of the sum of response error and penalty term to realize the filter update.The experimental results on VOT2018 data set show that the tracking accuracy is improved by 3.5%,the robustness is improved by 4.7%,and the average expected coverage is improved by 3.1%;the tracking speed is improved by 24.8% and 60.0% respectively with or without GPU.