Research On Object Tracking Algorithm Based On Online Learning Model

In recent years,the rapid development of computer hardware devices has given rise to the wide application of the image acquisition devices in almost every aspect of people’s lives.As a result,the object tracking technology which aims at researching the dynamic target images has received enormous attention of researchers both home and abroad in the field of computer vision.Being an interdisciplinary and comprehensive technology,the object tracking technology integrates many frontier research theories such as the digital image processing,pattern recognition,machine learning,mathematical statistics,biology and psychology etc.As one of the core technologies in the field of computer vision,the object tracking technology has now integrated into all the aspects of people’s lives and has demonstrated vast potential for future application in such areas as intelligent video monitoring,human-computer interaction and the national defense construction.In the computer vision field,the trailing work of object recognition and the behavior analyses are based on the accuracy of target tracking.However,due to the complexity of object’s surroundings and the variability of object’s appearance,the attempt to develop a pervasive algorithm of object tracking remains quite challenging.Under such circumstances,this dissertation makes in-depth researches on current on-line object tracking algorithm and proposes a series of on-line target tracking algorithms to improve the tracking robustness in response to the drawbacks and shortcomings in the on-line modeling and learning process of the target appearance model.The main research content and work of this dissertation are as follows:As a newly-proposed detection tracking algorithm,the compressed sensing tracking algorithm first updates the appearance model in the compressed domain and then locates the target with the help of the model.Nevertheless,the accuracy of the single classifier adopted in the compressed sensing tracking algorithm would be increasingly reduced provided that the target position area appears to drift.Apart from that,the compressed sensing tracking algorithm adopts constant update classifiers on learning rate whose target feature would be replaced by occlusion feature and losing the target eventually after some serious occlusion for a relatively long period of time.To solve the problems above,this dissertation puts forward a compressed sensing tracking algorithm based on mixed classifiers’ decision.In the process of tracking,this algorithm locates the target via mixed classifiers’ decision model and accordingly comes up with a strategy to update the appearance model in light of the dynamic learning rate,which greatly enhanced the tracking algorithm’s efficiency in coping with such interference factor as occlusion and background clutter.Featured as a kind of generating scheme,the tracking algorithm of L1 sparse representation can effectively adapt to the changes in targets’ appearances but at the cost of losing target’s background information.The compressed sensing tracking,in comparison,is a pattern of discrimination able to make full use of the target’s background information but in short of an effective mechanism to assess the findings in the process of tracking.For the purpose of integrating the strengths of both tracking algorithms,this dissertation put forward a compressed sensing tracking algorithm with the fusion of L1 sparse representation detection.The novelty of this algorithm is that the L1 sparse representation would verify make feedback on the tracking findings derived from the classifiers,which greatly assist the selection of classifiers at different levels to track object.In the meantime,in response to the template-updating issue in the dictionary of sparse representation occurred in the process of tracking by classifiers of different stages,a two-stage template updating strategy is adopted to maintain the most representative template in the dictionary,which therefore enables the integrated tracking algorithm to achieve better robustness under complex environment.For the past few years,the tracking algorithm on the basis of kernelized correlation filter(KCF)surpasses the majority of previous tracking algorithms in terms of precision,success rate and speed.However,when there is drastic changes in the object’s appearances,the algorithm of KCF,which updates the filter each frame,is easily subject to target drifting in the process of tracking.In addition,resulted from the filter used in the KCF tracking algorithm to track targets,the filter model is likely to lose the actual appearance of the target under the long-term interference of occlusion and eventually leads to a failure in relocating the target once it gets lost.To solve the problems mentioned above,a tracking algorithm based on multi-stage correlation filter is put forward in this dissertation which innovates in the following aspects.First,a multi-phase filter model of learning including the global phase,consistent phase as well as initial phase is proposed to perform parallel tracking to the object.On this basis,a "pure" training sample will be selected in ways of forward and backward tracking to update the filter at the consistent stage which helps forge a complementary relationship among the three appearance models corresponding to the three filter models.It is shown by the experiments that the multi-stage filter model of learning is greatly helpful in solving not only the drifting problems in the KCF algorithm but also capable of relocating the target within a local domain.