Algorithm Study On Targettracking Via Object Proposalsand Kernelized Correlation Filters

Object tracking is one of the fundamental and challenging research topics in computer vision,which has a variety of practical applications in the areas such as video surveillance,robot navigation and human-computer interaction.Given the initial state of a generic target in the first frame,the goal of tracking is to predict states of the target in the incoming frames.Recent years,with the rapid development of the state-of-the-art deep learning,correlation filter,object proposal,and classic machine learning approaches,a considerable amount ofexcellent tracking algorithms have been proposed.However,designing a real-time,robust and accurate tracker remains a difficult problem due to the following key issues.(1)The dilemma of model complexity and real-time requirement.Because of the time-sensitive nature of tracking,existing trackersarehard to find a delicate balance between exploring more complex models and keeping their computationaldemands as low as possible.(2)The drifting problem caused bymodel update.A tracker is prone to drift because it typically updatesusing corrupted and inadequate online training samples in a self-learning manner.(3)The tracking failure due to heavy occlusion or object moving out-of-view.The tracker may lose the targetand can't recover in an efficient way.It is a common practice to exhaustive search in a sliding window fashion.To address the above issues,current top performing trackers explore diverse solutions from different aspects.The Kernelized Correlation Filters(KCF)based trackerhasaround more increasing attention and performs well against others trackers,because it exploits the property of circulant matrix to analytically incorporate thousands of virtual samples without increasingmodelcomplexity explicitly,while using kernel method and Fast Fourier Transform to speed up the running time of computational procedure.Several effective improvement methods are proposed on the basis of previous research.The Main research contents are summarized as follows:(1)Different from traditional KCF which only leverage one layer and a single kernel at a time,wepropose aonline collaborative training of multi-layer multi-kernel correlation filters(MLMKCF)based tracker which fully takes advantage of hierarchical structure to enrich the object representation.The shallow structure of the basic KCF is firstly naturally yet effectively extended to a deep structure with multi-layer multi-kernel learning.Meanwhile,an efficient gradient descent method is derived to solve the weight distribution of thekernels.Then,to alleviate the self-training problem,we construct two MLMKCF with complementary types of featuresof object appearance,which collaborativelyupdate and dynamicallyfuse by multi-view correlation response.Finally,we carry out experiments on an object tracking benchmark(OTB-50),which demonstrate that the proposed tracker performs well in terms of accuracy and robustness.(2)To cope with the other problem that the KCF lack of detection failure,because objects suffers from significant appearance variation due to heavy occlusion,distracting regions and out of view,we proposed robust visual tracking via detectionproposals combined with kernelized correlation filter(MKCFDP),which replace the original sliding windows based on the objectiveness measure.The task of model-free object tracking is decomposed into tracking and detectiontwo modules.The tracker follows the object by translation and scale variations.The translation is efficiently predicted by modeling multi-layer multi-kernel correlation filter and the scale is independently estimated by searching the targetpyramidappearance.The detectorre-localizes the object position byan multi-experts ensemble,which is constituted by a base Support Vector Machine(SVM)classifier and its historical snapshots.The detectoris activatedwhen the peak-to-sideloberatio(PSR)belows a threshold,and the best classifier measured by a minimum entropy criterion is trained to re-rank proposals,which are generated by the most promising proposal Edge boxes.In addition,concentrating on image regionswhere edge information is eminent allows efficientselection of more object-like proposals,which is more efficient than exhaustive search scheme.Experimental results on OTB-50 also show that the proposed algorithm performs well in terms of efficiency,accuracy,and robustness.