Single Target Tracking Via Local Patches And Contextual Information In Complex Scene

Visual object tracking is one of the most important research topics in computer vision,and has been widely used in video surveillance,human-computer interaction,intelligent robot,augmented reality,medical image analysis,and etc.In general,visual object tracking is deemed as estimation and judgment of the target state.The goal of visual object tracking is to estimate or discriminate the target state in the consecutive image frames by given its initial target state(such as location,size and other information).In the past decades,visual object tracking has made remarkable progress.Particularly,it can achieve fairly high performance in the constrained conditions or relatively simple environment,such as the tracking of rigid objects in a static scene.However,due to the complexity of the target and the background,it is still a challenging issue to achieve robust and accurate target tracking in real world.In the process of tracking,tracking performance will be affected by various factors,such as part or full occlusion,illumination changes,in-plane rotation,out-plane rotation,background clutter,scale change and complex movement.These complex factors will lead to significant changes in the target's appearance.The existing tracking methods still cannot effectively solve the problems caused by these complex factors.In the process of visual object tracking,the representation often plays a decisive role on the tracking result and performance.In addition,the environmental factors of the target also have an important influence in the tracking process.Particularly,in the complex environment,target feature and contextual information of the target can provide important evidences for tracking.Therefore,it is one of the important ways to improve the tracking performance by effectively using target feature and contextual information.Based on in-depth analysis on the working mechanism of visual object tracking method,relevant work is carried out from the perspective of constructing robust appearance model.In this thesis,some new approaches are developed based on the local feature information and contextual information of the target.Our work and innovation could be summarized as:(1)To address the tracking issue when the target undergoes rapid and significant appearance changes,a novel tracking method is proposed.The object's appearance is represented by a set of local patches with inherent spatial geometric constraint relationship.It probabilistically adapts to the object's appearance changes by removing and adding local patches.Locations of new patches are determined by the global color property,and it can improve the traditional algorithms that the appearance model cannot be updated in time during tracking.The reason is that the proposed method is flexible to appearance changes by using local patches.At the same time,the global target model is constructed by color feature,which provides an effective priori information for the updating of local patches.As a result,it provides a more reliable and flexible updating manner for appearance model.Experimental results show that the proposed algorithm performs well in many cases,which has high accuracy in object tracking with drastically changes.(2)To achieve robust tracking in complex environment,a new visual tracking approach by using the local apparence model and the contextual information of the target is presented.In the tracking process,the object is represented by using the local patches and the context of the target.First,we decompose the bounding box of a target object into multiple patches,which are described by intensity and gradient histograms.Then,the likelihood defined as the weighted sum of reliability and stability indices is applied to evaluate the robustness of the patches in the tracking process.The method can effectively preserve the internal spatial structure information of the target by using parts-based representation mode,which can quickly adapt to the local appearance changes in the complex environment.Second,to suppress the drifts in the tracking procedure,we represent the target by using the contextual information which including foreground and background.By fusing the background information in the appearance model,the drifts caused by background blur and noise can be suppressed effectively.Experimental results demonstrate that the proposed approach can effectively deal with the complex conditions such as occlusion,fast motion,background blur,and etc.(3)To effectively represent the target object,a hierarchical model is presented in this thesis.Furthermore,a robust visual object tracking method based on hierarchical model is constructed under the Bayesian framework.In the tracking process,the proposed approach represents the target at two levels,i.e.,the local and the global levels.At the local level,a set of local patches are used to represent the target so as to adapt to the local appearance changes caused by occlusion,deformation,and etc.At the global level,the target is represented by the contextual information which including the foreground and background.As this,the drifts can be effectively suppressed which caused by background clutter,motion blur,and etc.Based on joint representation mode including the local and global layers,this method can effectively improve the accuracy of the target representation,and can effectively suppress the influence of various complex situation.Experimental results demonstrate that our method is very effective and performs favorably in comparison to the state-of-the-art trackers in terms of efficiency,accuracy and robustness,and it can achieved a high tracking performance in a variety of complex environments.(4)To achieve robust and accurate visual tracking,correlation filter is introduced into the tracking framework in this thesis.We propose a novel collaborative tracking method based on the correlation filter,local patches and the contextual information of the target.In order to improve the tracking accuracy,a coarse-to-fine tracking strategy is adopted.In the coarse tracking phase,the target is represented by a set of local patches,and each patch conduct the tracking task based correlation filter independently.Therefore,the object location is determined by the tracking result of each weighted patches.In the fine tracking phase,to suppress the effect of background noise,the target is represented by utilizing double bounding boxes which corresponding to the foreground and background,respectively.Then,a fine search strategy under the Bayesian inference framework is adopted to achieve an accurate location.In addition,to attain accurate tracking performance in the long-term tracking process,we discriminate whether a patch is under heavy occlusion by using its confidence map.In the process of tracking,by combining the generative model and the discriminative model,the algorithm has complementary advantages over the two models.In addition,the proposed algorithm combines the local tracking and the global tracking methods.Extensive experimental results on a set of sequences and benchmark dataset show that the proposed approach performs much better than the existing state-of-the-art methods do in terms of efficiency,accuracy and robustness.