Research Of Object Tracking Algorithm Based On Particle Filter And Sparse Representation

In the trend of informationization,some key techniques,such as computer vision and artificial intelligence,have been developing at a high speed.Using the ability of computers to replace the function of human eyes and brains has become one of the hottest topics in computer vision.Among them,object tracking has been widely used in both civil and military fields,i.e.monitoring and automatic positioning in battlefields.Meanwhile,many researchers have proposed lots of advanced algorithms inspired by the research difficulty of target tracking.Those algorithms can be divided into discriminative and generative models based on modeling methods.But so far,it is still a challenging task to achieve rapid and robust tracking with limited computational resources while dealing with changes in object poses and sizes,illumination and partial occlusion.As the both sides of discriminative and generative models,a new cascade-structured object algorithm that based on the support vector machine and sparse representation is proposed under the particle filter framework.The classification ability of support vector machine and the modeling capability of sparse representation are utilized into the collaborative models.The main research work includes the following three aspects:Firstly,the support vector machine based discriminative sub model is designed to alleviate the computing complexity of the numbers of particles.The initial classifier and the updated classifier are combined by the geometric meaning of confidence,the samples' selection and update methods are designed for keeping the initial target information and adapting to the state change simultaneously.The classification validity of the discriminant model is proved by experiments,in other words,the unreliable candidate particles can be selected out by the model.Secondly,in the generative section,the local based sparse generative model is improved accordingly.When solving the sparse equation,the two main algorithms,LASSO and OMP,are compared in order to find a compromise between the precision and the speed of the solution.And the entropy is proposed as the similarity measurement and its practicability in the scene of single inference factors is proved by experiments.Thirdly,the tracking evaluation platform,VOT(visual object tracking challenge),is used to analyze the performance of the proposed tracking algorithm.Both qualitative and quantitative evaluations on VOT demonstrate that the proposed algorithm has faster speed than the traditional collaborative model and performs favorably in the scenes that suffer occlusion,illumination variation and rotation,outperforms other state-of-the-art algorithms.