Research On Robust Object Tracking Based On Sparse Analysis

Object tracking,which embraces image processing,pattern recognition,machine learning,probability and matrix theory,is one of the popular research areas in computer vision.With the development of computer and video camera equipment technology,object tracking is widely applied in intelligent video surveillance,human computer interaction,traffic monitoring and military industry.However,it is still a challenging problem to develop a robust and real-time tracker for complex and dynamic scenes due to appearance changes caused by varying illumination,pose variation,occlusion,scale variation and complex background.In this dissertation,the research of sparse analysis based single object and multiple object tracking algorithm has been deeply done by collecting,collating and analyzing abundant domestic and foreign relevant papers and literatures.The major contributions of this dissertation are as follows:1.An illumination invariant color classification basing on sparse analysis model is proposed.As the fact that color feature is directly affected by environment illumination,it is still a challenging problem to classify the color feature of same target under different illuminations,which is also a key issue to develop an illumination invariant tracker.Motivated by the observation that the RGB value distributions of the same color under different illuminations are located in an identical hyperline,the color classification is formulated as a hyperline clustering problem via a sparse component analysis model.Experiments demonstrate the outstanding illumination invariant performance and robustness of the proposed algorithm as compared to existing clustering algorithms.2.A hyperline clustering baesed color object tracking robust to illumination change is presented.Color histogram is a simple yet effective statistic based feature description.Trackers relying on the strength of color histogram are robust to object deformation.However,they are sensitive to illumination changes.To overcome this limitation,a novel hyperline clustering based-discriminant model,which is able to distinguish the object from its surrounding background,is proposed in this chapter.Furthermore,an anchor based scale estimation is presented to cope with shape deformation and scale variation.In order to speed up the tracker,only the hyperline of background is updated during tracking.Numerous experiments demonstrate that the proposed approach achieve favorable performance,especially in the illumination variant and shape deformation challenging situations.3.A real-time robust object tracking via fusing LI APG and spares histogram is presented.Spare representation,which has a capacity of high reconstruction accuracy,is widely applied in object tracking.Thought spare representation based LI APG(Accelerated Proximal Gradient)tracker is running in real-time and robust to illumination variation and partial occlusion,it fails in some challenging situation,like scale changes and heavy occlusion.To strength the robust of L1APG,a real-time and robust object tracking algorithm fusing L1APG and spares histogram matching is proposed.The sparse histogram likelihood between sample and temple is calculated to weight the observation likelihood of LI APG sample,which takes local appearance information into consideration.And an adaptive update scheme is presented to learn the appearance changes of object.Thus,the proposed algorithm is more robust to scale changes and heavy occlusion than LI APG.Experiments on variant videos and qualitative and quantitative analyses show that the proposed algorithm not only increase more than 90%accuracy of LI APG but also more robust to other spares representation based trackers.4.A K-Matching Pursuit method for fast object tracking is proposed.Sparse Collaborative Model(SCM)tracker,which integrates discriminative model and generative model seamlessly,is robust to various challenges.However,it involves a lot of manipulations of sparse representation so that it is computationally expensive,which leads to low frame-rate.To solve this problem,a K-Matching Pursuit(K-MP)tracker is proposed to speed up the sparse representations in SCM.K-MP reconstructs a K-sparse signal by selecting K bases of a dictionary via l2-minimization.We mathematically demonstrate K-MP is able to successfully reconstruct a K-sparse signal.And,the tracking experiments verify the higher computational efficiency and tracking accuracy of the proposed K-MP tracker than that of the LASSO based SCM tracker.5.In order to solve the problem that trackers fail to detect and re-identify object,a real-time multi-object tracking based on mobile camera network is proposed by integrating Fast Compressive Tracking(FCT)with a real-time detector.The detector handles the occlusions and entering/exiting of targets whilst tracker updates the appearance features of targets.An adaptive size random patch sampling is presented to alleviate the influence of background and to normalize the size of targets.In order to keep a consensus on the 1D of the targets with the network,a distributed database where ID and appearance of targets are updated over time and shared among all cameras is introduced.Experiments on various benchmark dataset demonstrate the better performance of the proposed approach than other multi-object tracking algorithms.