Research On Object Tracking Algorithm Based On Linear Representation And Graph Model

Object tracking has always been a critical yet challenging research topic in the field of computer vision,it involves many aspect knowledge including image process-ing,pattern recognition,artificial intelligence,computer technology and so on.with the rapid growth of the computer technology and automatic sensor techniques,object tracking has attracted many researchers' interest,which has a wide range of applica-tion in behavior analysis,visual surveillance,human-computer interaction,visual nav-igation of robots,medical diagnose,military guidance and so on.The task of object tracking is to enable the computer to imitate the motion sensibility of human vision,so it can distinguish and detect the motion object from complex scene.Although there has been significant progress for object tracking in the past decades,however,devel-oping a robust,real-time and accurate tracking algorithm to meet actual task is still a challenging problem due to many challenging factors(eg.illumination change,pose change,scale,occlusion,rotation,motion blur,background cluster).To solve these above problems,some improved tracking methods are proposed af-ter thorough analysis many classical algorithms,these improved methods can improve the accurateness and robustness.This research works and contributions are listed as follows:1.Object tracking algorithm based on generative methods always need to con-struct an object appearance model and design an effective searching algorithm,due to appearance change and patrial occlusion,it is very difficult to track object accurately.To avoid to obtain an not robust searching algorithm and inspired by the successful ap-plication of manifold ranking algorithm in the field of image retrieval,a novel tracking algorithm based on manifold ranking algorithm is proposed.For tracking,tracked re-sults are taken as labeled nodes while candidate samples are taken as unlabeled nodes.The goal of tracking is to search the unlabeled sample that is the most relevant to the existing labeled nodes.Therefore,visual tracking is regarded as a ranking problem in which the relevance between an object appearance model and candidate samples is predicted by the manifold ranking algorithm,so the candidate with the largest rank-ing score is regarded as tracking result.In order to reduce computation complexity,an efficient manifold ranking algorithm is used for constructing graph model.Mean-while,we adopt non-adaptive random projection to preserve the structure of original image space,and a very sparse measurement matrix is used to efficiently extract low-dimensional compressive features for object representation.Furthermore,a support set is constructed to add spatial context information,which is used to exploit the spatial layout relationship between object and background,so our algorithm is more robust to background clutters and occlusion.Experimental results on some challenging video sequences verify the effectiveness of robustness of the proposed tracking algorithm.2.A novel tracking algorithm based on a weighted subspace reconstruction error is proposed in this thesis.The proposed tracking algorithm defines the discriminative weights in order to strength its distinguishing ability.The discriminative weights are defined by minimizing the reconstruction error using a positive dictionary while maximizing the reconstruction error using a negative dictionary,respectively.Based on the subspace reconstruction error,the thesis presents a simple yet efficient obser-vation likelihood function to combine the discriminative weights and the subspace re-construction error.The observation likelihood function not only distinguish object from complex background effectively,but also it can describe appearance change ef-fectively.Furthermore,in order to void object model degradation and tracking drift by inappropriate updating,the thesis propose a new valuation method based on a forward-backward tracking criterion to decide whether to update the appearance model or not.Experiments on many challenging video sequences have demonstrated the accuracy and robustness of the proposed method.3.The thesis proposes and effective visual target tracking algorithm based on the joint optimization of the online learning based combined spatial-temporal models.For tracking,the object can be linearly represented by object appearance model,while the object and its' background should meet constraint relationship,so temporal and spatial information are critical for tracking.The thesis firstly construct the temporal ap-pearance model based on PCA subspace method,this model can cope with appearance change.Then we use sparse representation to construct the spatial constraint model,which can exploit the relationship between the target and its neighbours.In order to reduce computational complexity and consider distance between input sample vector and dictionary basis vector,a K-nearest Local Smooth Algorithm(KLSA)is presented to describe the spatial state model.Further,a customized Accelerated Proximal Gra-dient(APG)method is implemented for iteratively obtaining an optimal solution in KLSA.Our method effectively fuses the advantages of a temporal appearance model with a spatial constraint model,to solve the weights of candidates to represent the optimal state estimate,then we can obtain tracking result based on particle filter frame-work.Furthermore,the thesis presents a reasonable model update method to improve the robustness of proposed algorithm.Finally,to compare with many state-of-the art tracking methods on many challenging videos,the proposed tracking algorithm obtains more robust tracking performance.4.The thesis presents a graph regularized and locality-constrained coding frame-work for robust object tracking.In the proposed algorithm,locality-constrained coding can ensure a sample with bigger coefficient that is closer to dictionary;in addition,our method ensure similar samples have similar coefficient by graph regularized con-straint,and the graph Laplacian as a smooth operator can ensure representations vary smoothly along the geodesics of the data manifold.Representation dictionary is it-eratively obtained by graph regularized locality-constrained coding algorithm,which can effectively describe the appearance change.After obtaining the representation dic-tionary,the thesis also presents an efficient observation likelihood function based on reconstruction error using dictionary.Furthermore,the thesis proposes a quick candi-date selection scheme.More candidates are sampled based on particle filter,and then this scheme can select a little of candidate to reduce computation complexity.Exper-iments on many challenging video sequences have illustrated that the effectiveness of the proposed tracking algorithm.