Research On Correlation Filter Based Object Tracking Methods And Key Technology In Complicated Conditions

As a core research in computer vision,object tracking has a wide range of application value in video surveillance,human computer interaction,military reconnaissance,precision,intelligent transportation,augmented reality,and medical diagnostics.In recent years,with the development of the artificial intelligence,the research on target tracking technology have made significant progress at both domestic and foreign.However,actual application scenario contains a variety of interference at the same time,such as background clutter,illumination change,occlusion,deformation,the target rotation,scale changes,motion blur and rapid movement,etc.The existing tracking algorithms are difficult to meet the realistic demand.Therefore,it has important practical value and research significance to study a feasible and robust tracking method.On the basis of the investigation analysis of the target tracking methods from both domestic and foreign,this dissertation does a deep research on object tracking method based correlation filter tracking framework,to solve some existing problems of correlation filterbased tracking in some complicated situations.The main innovations of this dissertation are summarized as follows:1.A novel object tracking algorithm based on online learning the topological structure of the context is proposed.Existing most object tracking algorithm focus on the research on the object appearance model,the purpose is to construct an appearance model with more generalization ability.However,when the scene has some features similar to the object appearance,the better generalization ability model,the more vulnerable to interference.This dissertation excavates the potential topological relations between object and similar objects,and propose to improve the discriminative power of the tracker by the learned the topological structure of the context.The nodes of the topology graph are denoted by the appearance models of similar object and real object,and the edges are represented by the directed link between the object and similar object.Then,to construct a joint learning model to online and jointly train the nodes and edge of topological graph.In addition,considering similar object is not inevitable,the dissertation introduce an interference weight into the model to reflect the interference power from similar object.Experimental results show that the proposed method can effectively alleviate the interference from similar objects.2.A consistently sampled correlation filters with space anisotropic regularization is proposed for visual tracking.The performance of correlation filter-based trackers benefits from dense sampling,but the cyclic sampling introduces the circular boundary effects into every sample.The boundary effects will become more serious when the sampling area is expanded,which limits the discriminative power of the model.In addition,most of the existing correlation filter-based trackers without considering the reliability of the sample,if the contaminated training samples are introduced into the training,performance of the tracker will become degradation.This dissertation proposes to construct and study a consistently sampled correlation filter with space anisotropic regularization to solve these two problems simultaneously.Constructing a spatiotemporally consistent sample strategy to alleviate the redundancies in training samples caused by the cyclical shifts,to eliminate the inconsistencies between training samples and detection samples,and introducing space anisotropic regularization to constrain the correlation filter for alleviating drift caused by occlusion.Experimental results have demonstrated that the proposed algorithm performs well in a wide range of tracking scenarios,e.g.,background clutter,occlusion and motion blur.3.An online adaptive complementation-tracker is proposed.Target tracking algorithm based on correlation filter-based tracker is sensitive to structural information naturally.The attribute is helpful to locate the rigid target,but is not suitable for dealing with deformation.This paper introduces scene perception into the tracking,and constructs an adaptive complementary tracking algorithm to alleviate the deformation.Two sub-trackers from complementary model,one of them focus on the structure features of the scene while other prefer non-structure features.Using the history tracking results to train each sub-tracker,and through the model reliability weight to construct a linear combination from multiple historical regression models to generate a joint learning model.Jointly learning the subtrackers and their reliability weights by the joint learning model.Then,using the statistical properties of these reliability weights to build adaptive complementary weights.Experimental results show that the proposed method has achieved favorable tracking performance comparing with the most of the existing correlation filter-based tracking algorithm.