Study On Object Tracking Algorithm Based On Sparse Representation And Nonnegative Matrix Factorization

Object tracking is one of the hot issues in the field of computer vision and pattern recognition.It has been widely used in such applications as intelligence navigation,automatic surveillance,military defense and human-computer interaction.Although object tracking has got quite great progress after decades of development,they usually do not completely and effectively deal with object variation due to noise,rotation,partial occlusion,motion blur,and illumination or pose variation.Recently,the generation model-based object tracking algorithm is got a lot of focus because of its robust tracking performance via adaptively adjusting itself according to the changes of target appearance.The aim of this thesis is to study on the generation model-based object tracking algorithm.Inspired by the sparse and non-negative property of video sequences,the sparse representation theory and non-negative matrix factorization are adopt into the generation model to represent and update the observation object model.The proposed algorithms are used to conduct object tracking under various complex backgrounds.The major work of this thesis are as follows:(1)This thesis proposes a object tracking algorithm with template prior probability and sparse representation.In the L1-based sparse represent object tracking algorithm,the superior target templates are usually updated or deleted in the subsequent object tracking process,resulting in decreasing tracking performance.To tackle this problem,the cumulative selection method is used to select high-probability template as the key template.In order to highlight the role of the key template,a new template update strategy is proposed,where the importance of templates is introduced into the regularization model to further detect the object from candidates.In addition,the accelerated proximal gradient approach is used to solve the aforementioned regularization model to improve the real-time tracking performance.Experimental results of object tracking under several common complex backgrounds show that the proposed algorithm is effective and robust,especially in the case of occlusion of targets by similar objects.(2)This thesis proposes an incremental projection non-negative matrix factorization-based object tracking algorithm with both smooth and sparse constraints.Firstly,a part-based subspace is learnt incrementally to represent the target.The sparse constraint is used to improve the ability to deal with occlusion,and time smoothing constraint is adopt to improve the stability of the algorithm.The optimization of the observation model is solved via multiplicative iterative updating rule to decrease the computational complexity.In addition,based on the oberservation of the sparsity of coding coefficients,a new observation likelihood function is proposed.Experimental results on various video sequences show that,compared with the state-of-the-art tracking methods,the proposed algorithm can deal with large occlusion or scale variation of the object.