Research On The Application Of Sparse Representation In Object Tracking

In recent years,the rapid development of target tracking technology has been widely used in various fields.The target tracking technology can be attributed as the prediction of target motion trajectory tracking technology.This prediction or tracking depends on the accuracy of the tracking algorithm itself.At present,there are two kinds of tracking algorithms,one is online learning method,the specific performance for the generated method and discriminated method,the other is based on the sparse representation.In the framework of sparse representation,two kinds of sparse target tracking algorithm are proposed in this thesis,and the experimental results show that the two tracking algorithms have shown good robustness and accuracy.Firstly,the Gauss noise and Laplace noise are considered in this thesis.An adaptive sparse representation of target tracking algorithm was proposed.This algorithm not only fully consider the effect of noise,but also propose new template selection and update mechanism,namely online incremental learning and K-means combination method of template update.The new template updating method can effectively weaken the tracking drift phenomenon and gives the template update method,the distribution of noise and the combination of the two effects on tracking results respectively.Experimental results show the robustness and accuracy of the proposed algorithm has stronger robustness and accuracy.Secondly,in order to improve the real-time tracking system,the adaptive pL sparse target object algorithm is proposed in this thesis.The algorithm with the former proposed algorithm,it also considers the effect of noise distribution and uses the template update method with online incremental learning combined with K-means method.The only difference is that the model considered particles are independent each other.Each particle can be computed respectively through sparse model,and LASSO algorithm is also used in the model.The optimal tracking results are calculated according to the framework of particle filtering theory and maximum a posteriori probability.The numerical results show that the proposed algorithm is better than the state-of-the-art tracking algorithm,it has better tracking effect for performance higher accuracy and better real-time.Finally,direction of future research and the problems is given in this thesis,and work for the next step is prospected.