Research On Object Tracking Algorithm Based On Sparse Representation

At present,the target tracking algorithm is affected by occlusion,the changes of scale and pose,rotation and other factors,the real-time and accuracy of tracking is sometimes difficult to meet the practical application needs.In recent years,the particle filter algorithm and sparse representation theory have been widely applied in target tracking field,particle filter algorithm can effectively fit the motion state of the target,and sparse representation can highlight the main features and key characteristics in the image.In the framework of particle filter,combining the sparse representation theory,we propose a target tracking algorithm based on sparse representation.The main research work of this thesis is as follows:1.Aim at the fact that tracking failure of particle filter often happens in the cases of low discriminative abilities of the observed features and disperse distributions of the sampled particles,we introduce a multifeature uncertainty measurement function to adaptively adjust the relative contributions of different features.Meanwhile,we construct a self-adaptive feature fusion strategy to overcome the shortcomings of product and sum fusion ones,this strategy effectively sharpens the distribution of the fused posterior.Finally,we combine HSV and HOG features into an adaptive fusion framework,and propose a target tracking algorithm based on adaptive fusion of HSV and HOG feature.Simulation results show that the proposed algorithm is more stable and robust than the target tracking algorithm of single feature and traditional feature fusion when the target object in variation of pose,scale and rotation.2.Aiming at the problem of the sparse representation only represents the independent image blocks,and ignores the importance of the correlation of visual information When the object is represented,and in order to further improve the tracking accuracy of the algorithm.In this thesis,a new target tracking algorithm with collaborative sparse representation of multifeature is proposed,which is based on the particle filter framework and sparse representation.Firstly,in the construction of the target template,both the fusion feature and HAAR features are used to describe the target,the appearance model of the tracked target is modeled by instantaneous and stable appearance features simultaneously.Then,a two-stage sparse-coded method which takes the spatial neighborhood information of the imagepatch and the computation burden into consideration is used to construct the tracking likelihood function of transient and stable appearance.Finally,the reliability of each tracker is measured by the tracking likelihood function,and the most reliable tracker is obtained by a well established particle filter framework.The templates library are incrementally updated based on the current tracking results.Experimental results show that the proposed method is robust to pose variation,scale change and partial occlusion.