Study On Moving Object Tracking Algorithm In Surveillance Video Sequences

With the rapid development of electronic information technology,visual tracking has been fully integrated into the military,industrial production and daily life,such as the military UAVs,the industrial robots and the security monitoring.With the development of visual target tracking for decades,many excellent algorithms have been developed.While these algorithms still exist many problems in practical application,especially when the following phenomenon occurs: illumination change,occlusion,deformation and similar targets and so on.This paper mainly study the problems of occlusion,deformation and low positioning accuracy in visual tracking.The main results of this paper are as follows:?1?A kernel collaborative representation tracking algorithm is proposed.In the traditional template matching algorithm,the candidate needs to be constructed by a series of linear combination of the target template,but this model has some limitations on the nonlinear problem.In order to overcome the limitation of the traditional algorithm,a kernel cooperative representation method based on the particle filter framework is proposed,which can effectively track the target.The algorithm uses the kernel function to map the target template and the candidate target in high dimension,and measuring the similarity between the candidates and templates in the high dimensional space.At the same time,the l₂ norm constraint is introduced to reduce the computational complexity.?2?A tracking algorithm based on the confidence value reconstruction is proposed.Motion deformation and occlusion are two common problems in tracking technology,and also the most difficult problem.In order to reduce the impact of these problems,we firstly divided the template and the candidate into several parts,and use the patches of the templates to construct the dictionary.Then,the local patches of the candidates can be classified by the probabilistic representation,and their classification probabilities have been further acquired.The confidence value of each candidate is constructed by using the classification probabilities of its local patches.Finally,the confidence values of the candidates are reconstructed by using the block classification probability,and the candidate with the greatest confidence value is the target of the new frame.?3?A tracking algorithm based on fusion of two confidence maps.The kernel correlation filter tracking algorithm which estimates the position of the target by calculating the correlation between the filter template and the candidate has tracking error when the target is deformed.The local color histogram is used to improve the tracking accuracy of the tracking algorithm.The algorithm uses the local color histogram of the target to highlight the target in the search area,and obtains the confidence map of the target position by using the integral image method.Finally,a more accurate estimation of the target's position is obtained by fusing the confidence map of the kernel correlation filter tracking algorithm.