Research On Object Tracking Algorithms Based On Sparse Representation

Online object tracking is one of important branches of computer vision.It has great applications in motion analysis,activity recognition,visual surveillance,traffic monitoring and so on.Because of the changes of object and the environment during the tracking process,the appearance of the object is also changing,which produces some influences on the results of object tracking.Therefore,designing a robust object appearance model is a key problem.This dissertation focuses on visual object tracking based on sparse representation theory in the framework of the particle filter.With object appearance changes and object information redundancy problem to improve the tracking accuracy and robustness of tracking results.It proposes a robust object tracking using a sparse coadjutant appearance model,a robust object tracking using discriminative group-sparse representation,and a robust object tracking via feature selection and time-consistency sparse appearance model.The concrete contents are as follows.1.Robust object tracking using a sparse coadjutant appearance model.Firstly,we use the property that the global feature of an object can distinguish the object from the background well to design a sparse discriminative similarity model called discriminative score model.In the meanwhile,we use the property that the local feature information of an object can handle the changes of the target appearance to design a sparse generation model called similarity measurement model.Secondly,we integrate the discriminative score model and similarity measure model to generate a sparse coadjutant appearance model.Lastly,we design two update schemes that update the object templates of the discriminative score model and the template histogram of the similarity measurement model.The proposed object tracking algorithm considers the global and local feature informations of the object simultaneously.Experiments demonstrate that the proposed algorithm improves the accuracy and robustness of the tracking results.2.Robust object tracking using discriminative group sparse representation.There are a lot of redundant information during sampling the object template and candidate object,which leads to candidate object's discriminative features unfaithful.To overcome this problem,this paper proposes an object tracking algorithm using discriminative group sparse representation.Firstly,we build a multi-task group sparse representation model by introducing a group sparse penalty term as well as a sparse penalty term.Secondly,we use the alternating direction method of multipliers(ADMM)to solve the multi-task group sparse representation model to obtain the candidate objects' discriminative score.Lastly,we design an update scheme for the object template to improve the robustness of tracking algorithm.With group-sparse representation and ADMM,the proposed tracking algorithm makes each row of discriminative sparse similarity map more sparse and gets rid of the redundant information during sampling the candidate object,which guarantees the discriminative features of each candidate object more accurate to improve the accuracy and robustness of the algorithm.3.Robust object tracking based on feature selection and temporal consistency sparse appearance model.Firstly,sample some positive and negative templates as well as candidate targets,select their corresponding features according to the feature selection model,get rid of the redundant interferential information,and obtain the key feature information.Secondly,establish a multi-task sparse representation model that contains a temporal consistency regular term via the features of positive templates,negative templates,and candidate targets,which makes more candidates have the similar sparse representation with the previous tracking results.Thirdly,solve the multi-task sparse representation model to get the discriminative sparse similarity map,and obtain the discriminative score for each candidate target.Lastly,update the positive templates and the negative templates according to the tracking results.With the feature selection on the object templates and candidate objects,and the introduction of temporal consistency regular terms,the proposed algorithm improves the accuracy and robustness of the tracking results.