Research Of Object Tracking Based On Incremental Learning And Linear Representation Model

Online object tracking is one of the main core issues in the field of computer vision,which has extensive applied value in video surveillance,human-computer interaction,behavior analysis,and so on.Although online target tracking technology has a significant development in the past few decades,designing a robust online object tracking system is still a much challenge task for both the inner and external interferences including posture change,occlusion,illumination,shake of camera,and so on.This paper pays attention to the study of linear representation model based on the incremental learning,and designs robust online object tracking methods.The main tasks include:(1)This paper concludes and summarizes the research background and significance of online target tracking technology.Moreover,the research status,technical classification and research difficulties of object tracking in domestic and overseas are also briefly introduced in this paper.This paper summarizes the relative theories of online object tracking,and introduces the relative theories based on incremental visual tracking,including the framework of Bayesian lemma,the sampling based on affine parameters,the PCA(Principal Component Analysis),and the subspace incremental learning algorithm.(2)This paper shows an object tracking method based on incremental learning and collaborative representation.This method applies the PCA and the collaborative representation based on L2-norm regulation to learn the appearance model for object.Compared with the traditional incremental learning method using the linear combination of subspace basis and trivial templates for object representation,this paper directly models the residual between observed sample and reconstructed sample as Laplacian to reflect the interference of outliers,and alleviate the ambiguities of trivial templates that reconstruct both the background and foreground.Meanwhile,the proposed alternative and iterative algorithm within the framework of Augmented Lagrangian can be effectively applied for the minimization of objective function.The experimental results show that the proposed algorithm can achieve satisfying tracking precision without sacrificing large amount of calculation.(3)This paper shows an incremental learning object tracking method based on Lp-norm regulation.As the PCA subspace may import redundant background information when incremental update,we apply the Lp norm to regulate the target coefficient of subspace to achieve the purpose of sparse representation.Compared with traditional sparse representation mechanism,this paper proposes a generative sparse representation framework.In the condition that keeping same punishment to outliers,this framework can observe the tracking results with different sparsity for the same video sequence.Meanwhile,within the framework of APG(Accelerated Proximal Gradient),we also introduce the GST(Generalization of Soft-threshold)algorithm for the minimization of target coefficient.In order to demonstrate the effectiveness of the algorithm,this paper adopts the L0.5-norm as the example in multiple video sequences,and obtain high tracking precision.To observe the effect of excessive sparsity,in the context that all the parameters are fixed,this paper also observe the tracking results in these sequences with L0-norm regulation,and give some analysis and discussion for the experimental results.Finally,this paper also summarizes the main studied results,and also makes some outlook for the online tracking technology in future.