Research On The Combination Of Siamese Network And Correlation Filter In Visual Tracking

Visual tracking is an important branch in computer vision.Recently,relying on the rapid development of Artificial Intelligence and Big Data,it has been widely used in video monitoring,automatic driving,behavior recognition and other fields.The basic task of visual tracking is to locate the right target in an image sequence and mark it out.In the real scene of visual tracking,there are various factors which will seriously affect the accuracy of tracking results,such as target deformation,illumination variance,occlusion and so on.Therefore,how to design an efficient and robust tracker to deal with complex tracking environment is a challenging research topic.The current research direction of visual tracking can be divided into two parts:trackers based on siamese network and trackers based on correlation filter.These two kinds of trackers use different models to deal with the visual tracking problem,and they also have their own advantages and disadvantages.This dissertation studies these two kinds of trackers carefully,focusing on the combination of siamese network and correlation filter.Specifically,this dissertation is trying to improve and optimize correlation filter's tracking theory according to the characteristics of siamese network tracker,and then intergrate correlation filter theory into siamese network tracker to solve siamese network tracker's problems,such as the lack of target's prior information,the limitation of the features and the lack of self-adaptivity during tracking.Motivated by this idea,this dissertation mainly does the following three tasks:Firstly,this dissertation proposes a theory which can build a model of the target's prior information to improve the tracker's performance,and on the basis of this theory,a tracker named HKSiamfc(Histogram-Kalman Siam FC)is proposed.As the original siamese network is trained offline,so this leads its lack of enough prior information of the target,which will affect its tracking accuracy.The HKSiamfc tracker is proposed on the basis of original siamese network.This new tracker can use the technology from correlation filter theory and kalman filter to model the target's prior information of appearance and motion,respectively.So it can make good use of the target's prior information during the tracking task and shows robust performance.Secondly,this dissertation proposes a theory which can combine different trackers and make them work together,and on the basis of this theory,a tracker named MFCFSiam(Multiple-Features-Correlation-filter Siam FC)is proposed.Specifically,this dissertation studies the working mechanism of the original siamese network tracker and the correlation filter tracker,and summarizes their problems in the training mode,feature usage mode,search area and so on.We find that there are some complementary characteristics between these problems,that is to say,the shortcomings of one tracker can be compensated by the advantages of the other.The MFCFSiam is proposed on the basis of this discovery,it can realize the complementary cooperation between the two trackers.It regards the siamese network tracker and correlation filter tracker as its subtrackers and combines them in a parallel way.MFCFSiam also contains a component that can select the better tracking results between the sub-trackers.Therefore,in this framework,the siamese network tracker and correlation filter tracker can give full play to their advantages,make up for each other's shortcomings,and effectively realize the complementary tracking.Thirdly,this dissertation proposes a theory which can adaptively update the template of the original siamese network,and on the basis of this theory,a tracker named Apce Siam(Average Peak-to-Correlation Energy Siam FC)is proposed.Specifically,the original siamese network tracker lacks an self-adaptive updatemechanism for the target template.In order to solve this problem,this dissertation studies the concept of APCE(Average Peak-to-Correlation Energy)from correlation filter theory,and introduce it into siamese network tracker,and proposes a new tracker.This tracker contains a self-adaptive coefficient generated by APCE for the target template,which can be used to control the process of template update dynamically.In addition,a lot of experiments are conducted on several visual tracking dataset.The experimental results prove that the performance of the three trackers proposed in this dissertation has showed obvious improvements when compared with the baseline trackers.For example,on the OTB100 dataset,the three trackers we propose have achieved 7.2%,4.9% and 2.1% improvement in the success rate respectively,and 3.4%,4.9% and 2.1% improvement in the success rate respectively.And what's more,those three trackers also shows sufficient competitiveness when compared with other stateof-the-art trackers.