Object Tracking Research Based On Mean Shift Algorithm

In the various applications of computer vision and digital video processing, moving tartget tracking is an important and basic task.It fuses the advanced technology and research results on Image Processing, Pattern Recognition, Artificial Intelligence, Automatic Control and Computer Applications,etc. It has a very wide range of applications in the smart monitoring,human-computer interaction,image compression and national defence,ect.Mean Shift is an efficient pattern matching algorithm , it uses the nonparametric density estimation technology to climb to the density maxima along the direction of density gradient quickly and efficiently. Because it do not need to conduct exhaustive search, it has been successfully used in higher real-time demand on target tracking domain.This thesis mainly focuses on the moving object tracking based on Mean Shift algorithm under the complicated background. At first,it makes classification for the moving target tracking algorithm, and introduces the features seletion for tracking and the similarity measure functions, then it introduces the basic principle of the Mean Shift algorithm and proves its convergence in details. Finally the thesis focuses on raditional Mean Shift moving object tracking algorithm under the complicated background, the candidate target model is easy to be disturbed by the background information, besides tracking is usually failed with the fixed template. To solve the above problems, this thesis gives an improved Mean Shift moving object tracking algorithm, which puts the initial target model into the candidate target model, because the initial target model describes tracking target better, and the target model rarely contains background information, meanwhile the improved algorithm setting a threshold to judge the pixel is the target model or background information, thereby to diminish the nontarget's interference to the candidate target model, and gets the new weights which enlarges the possibility of the pixels is likely belonged to the target,thus it eliminates the influence from nontarget in target model for the tracking result. In view of the influence of fixed bandwidth, the method uses the adaptative kernel bandwidth to adapt the change of the target model, so then to increase the accuracy of the algorithm. The simulation result shows that the improved algorithm can effectively tracking objects, and can finely solve the problem from the tracking process under the complicated background.