Research On Object Tracking Based On Kernel Correlation Filter

Object tracking belongs to the field of video analysis and is widely applied in the field of life,traffic,medical treatment.However,due to complexity in reality,there remains a lot of challenge,such as scale and rotate variation,occlusion,deformation,similarity,etc.Based on the framework of kernel correlation filter,this thesis deeply analyzes and researches on scale variation,occlusion and fast deformation.The main research work of this thesis is as follows:Firstly,an effective scale estimation KCF tracker based on sparse feature is proposed.On the basis of KCF,the adjustable Gaussian window is introduced for better foreground and background separation,and correlation filter and sparse feature match are combined to estimate location and scale of target,which allow the correlation tracker react to scale changes.Furthermore,the scale estimation is interpolated into the KCF pipe but does not change it.Therefore,this method can be applied to any variant of naive CF tracker whether it is a KCF using Gaussian or polynomial kernel,or a DCF using linear kernel.Secondly,an anti-occlusion tracker based on KCF is proposed.The SVM classification is introduced for re-detection on the basis of KCF,and a double criteria is applied to decide when to update the model appropriately.In this algorithm,temporal context model based on kernel correlation is used for the preliminary location of target,and multimodal detection is applied for better translation if multiple peaks exists in the response map.While appearance model based on kernel correlation is trained by a patch whose size is less than bounding rectangle.When later regression score is lower than that of threshold Tr,the SVM classification will be transferred for re-detection and the result is chosen according to re-correlation score.Furthermore,former model updates each frame but later model and SVM classification update only when double criteria reaches.Thirdly,a KCF tracker based on mixed model is proposed.On the basis of template learning using HOG description,this tracker combines pixel-wise learning using histogram of color to maintain mixed model for adaptation of fast deformation and fast motion.To maintain high efficiency,two independent problems are solved by exploiting the inherent structure of template match and naive Bayesian.Due to the similarity of these two models in amplitude,they can be combined in linear by computing Gaussian fitting function respectivelyFinally,a multiple object tracing method based on background subtraction and KCF is proposed.Background subtraction can give noise-containing candidate object regions which should be analyzed and processed for high-quality potential objects.For tracking,appearance model using color naming assign KCF trackers for each candidate regions.In order to solve scale changes,fragment,and missing tracking,KCF trackers and background subtraction are used to create simple data association.Due to problems in different time for these two methods,the states of objects are determined by combinations of KCF and background subtraction.In this thesis,the performance of proposed algorithms are examined in open video sources,including quantitative analysis and qualitative analysis.Based on this,every proposed algorithm are respectively compared with several impressive ones.The results shows that the scale estimation based on sparse feature and adjustable Gaussian window can effectively do with scale changes in KCF;The anti-occlusion method based on SVM re-detection and update strategy can deal with occlusion in KCF;The algorithm based on mixed model can react to fast deformation of target in KCF.Furthermore,KCF can be extended to multiple object tracking.