Research On Object Tracking Via Non-negative Weighted Sparsity-based Collaborative Model

In recent years,target tracking has played a very important role in the field of computer vision,and has been a very important research topic.With the continuous development of target tracking technology,it plays a vital role in various practical applications.Therefore,a variety of target tracking methods came out,mainly including the tracking methods based on generative appearance model,the tracking methods based on discriminative appearance model,and the combination of them.In the last few years,the sparse representation has a good effect on difficult problems in target tracking such as occlusion.The object tracking algorithm via sparsity-based collaborative model(SCM)is robust and has good performances.The algorithm combines the sparsity-based discriminative model based on holistic templates with the sparsity-based generative model via local representations in the framework of particle filter,which takes advantages of both of them,and achieves good results in the target tracking problem.Non-negative sparse representation is based on sparse representation with non-negative constraints,which makes the objective function having a certain physical meaning.Because the obtained data and the extracted features in the process of target tracking have some non-negative properties,which are consistent with the premise of non-negative sparse representation,futhermore,the sparse generative model based on local histogram in SCM has a potential non-negative requirement on the coefficients.Therefore,we introduce non-negative constraints into SCM,and propose a target tracking algorithm which names non-negative sparsity-based collaborative model(N-SCM)in our first work.At the same time,considering that there is a certain correlation between the test samples and training samples,adding a reasonable weight to the coefficients of sparse representation can obtain a good result.Therefore,our second work is to add the corresponding weights to the sparse coefficients in SCM,and propose a target tracking algorithm which is named weighted sparsity-based collaborative model(W-SCM).Our third work is adding weights and non-negative constraints to SCM,and proposing non-negative weighted sparsity-based collaborative model(NW-SCM).The contents of this thesis are summarized as follows:(1)In order to improve tracking results of object tracking via sparsity-based collaborative model(SCM)and make better use of original data,we add non-negative constraints to SCM algorithm,and propose a target tracking algorithm named non-negative sparsity-based collaborative model(N-SCM).An iterative updating strategy is proposed to solve the sparse coefficients in the N-SCM algorithm,and the iterative convergence of the objective function.is proved.The experimental results show that compared with SCM algorithm,the performance of N-SCM algorithm is improved.(2)Considering that it is often ignored the correlation between the test samples and training samples in the previous sparse representation,which makes the sparse coefficients there is room for improvement,so we weight coefficient of sparse collaboration model and put forward the target tracking algorithm named weighted sparsity-based collaborative model(W-SCM).W-SCM algorithm uses an iterative algorithm to solve the global optimal solution of the objective function,and gives thetheoretical proof for the rationality of the algorithm.Experimental results show that the tracking effect of W-SCM algorithm is better than SCM algorithm.(3)Adding weights and non-negative constraints to SCM algorithm can improve the performance on object tracking respectively.So we not only add non-negative constraints,but also weight coefficients of sparse collaboration model and put forward the target tracking algorithm named non-negative weighted sparsity-based collaborative model(NW-SCM).We propose an iterative updating strategy to solve the sparse coefficients in the NW-SCM algorithm,and prove the iterative convergence of the objective function.The experimental results show that the performance of N-SCM algorithm is better than SCM algorithm,N-SCM algorithm and W-SCM algorithm on some video sequence.