Research On Target Tracking Methods Based On Dense Net And Kernel Correlation Filter

Target tracking is an important research branch in the field of computer vision,which is widely used in the fields of intelligent surveillance,intelligent transportation,unmanned vehicles and military technology.However,due to the complexity and diversity of the target itself and the surrounding environment,some random problems in the target tracking process such as background interference,occlusion,scale change,target deformation and others will seriously affect the accuracy and real time of target localization,therefore,the use of new techniques of information processing to improve the performance of target tracking has always been an important research direction of researchers’ attention.With the rise and development of deep learning in recent years,scholars have started to introduce deep learning techniques into the research of target tracking.The performance of existing trackers is improved by acquiring deep feature information of targets through deep learning networks.Inspired by the previous work of the group,this thesis investigates the target tracking strategies combining dense convolutional networks with the classical kernel correlation filtering framework,which are discussed in depth in three aspects,including target localization,anti-background interference and anti-partial occlusion,respectively.The main research results of the thesis are as follows.(1)A target tracking method(M-Dense Net-KCF)based on an improved dense network combined with kernel correlation filtering is investigated.In order to obtain features with stronger representation,the number of dense blocks is adjusted in the base Dense Net model and the bottleneck structure is added to construct a new Dense Net model.After that,the adjusted network model is trained on a large-scale dataset;then the trained network model is used to extract depth features,which are used as training samples to obtain the best correlation filter.Finally,the final location of the target is determined by searching for the maximum value in the response graph.This technical study not only improves the localization accuracy of target tracking,but also increases the computing efficiency.(2)A dense network correlation filter tracking method based on background and feature points is investigated.In order to improve the influence of background clutter on target tracking,the Kalman filter is used to predict the motion trajectory of the target in the image firstly,and the correlation filter is obtained by training the background information on the motion direction region to improve the ability to identify the target and the background.After that,FAST feature points are obtained in the motion direction region,and then the target center and scale are predicted based on the feature point information matched between this frame image and the first frame image.Finally,the target center estimated are weighted and fused by this method and M-Dense Net-KCF to further locate the target accurately.This technical study effectively improves the background interference problem and also improves the scale problem,which further improves the tracking performance.(3)An anti-occlusion dense network correlation filtering tracking method based on high confidence update is studied.Firstly,based on the obtained target predicted position,two discriminative criteria of high confidence update are used to determine whether to update the model.Secondly,the occlusion discriminative factor and response stability factor are introduced using the target context information to determine the occlusion.Finally,the occluded target is re-detected using a support vector machine detector to recover the target position.This technical study effectively improves the occlusion problem and greatly reduces the problem of model drift.In this thesis,the simulation experiments of the proposed algorithm are carried out on the OTB100 public dataset,and the qualitative and quantitative comparative analyses of the experimental results are conducted with the current excellent tracking algorithms,respectively.The experimental results show that the target tracking method based on the combination of improved dense network and kernel correlation filtering can effectively improve the accuracy of target localization;the dense network correlation filtering tracking method based on background and feature points can effectively solve the problem of background interference in the tracking process;the anti-obstruction dense network correlation filtering tracking method based on high confidence update can effectively determine the obscuration state of the target in the tracking process,and accurately update the model to determine the target location.