Research On Mean Shift Object Tracking With Fusion Of Depth Perspective

Video object detecting and tracking is an important subject in applied computervision systems, including image processing, pattern recognition, artificial intelligence,etc., it has a wide range of applications in both industrial and commercial arenas. Withmassive information rendered by images, object detecting and video tracking play anindispensable role in security surveillance, man-machine interaction, commercialadvertisement, sports matches, thus providing great convenience to the human daily life.In this research, depth cues are applied in Mean Shift object tracking algorithm,based on the characteristics of the depth cues, kernel bandwidth self-adaption andtemplate self-update in Mean Shift algorithm are investigated. Also, research on refiningthe color feature histogram and enhancing the robustness, decreasing the iterationnumbers by studying the shift coefficient in Mean Shift algorithm are conducted. Thecontents of this paper are organized as follows.A depth cues weighted histogram in Mean Shift algorithm is proposed to address thecolor histogram problem of interfering with varying illumination, and the solution issought by projecting the color feature space to the depth cues apace, hence the objectcolor feature is highlighted and algorithmic iteration number is reduced, resulting withthe more efficient and better tracking performance.The kernel bandwidth self-adjustment mechanism using depth cues is presented. Thecorrelation between kernel bandwidth and depth cues is established through deriving thefunctional relation of kernel bandwidth and depth quantities. Moreover, the targetlocation correction and template update are implemented by building the virtual templatebased on the overall depth distribution.The role of the shift coefficient in Mean Shift in object matching is also studied. Toidentify the most useful color information in the kernel window and enhance the colorfeature histogram, the determination of the kernel size through the shift coefficient areprobed with great care, in which the color feature is mapped with corresponding shiftcoefficient in a bid to update the template.In addition, a template updating algorithm on the basis of the foreground objectextraction utilizing the Graph Cut in Mean Shift algorithm is studied using the priorobject depth information to extract the object of interest. To check the template updatingconditions when template updating is needed, the reverse Mean Shift algorithm is adopted.The template updating is carried out provided the conditions are satisfied. The template updating based on the virtual template in Mean Shift algorithm ispresented with focus on the object’s appearance and shape information, which can beutilized to update the color feature information. The template updating based on shiftcoefficients places extra emphasis on extraction of object’s color feature information,meanwhile separates the background information and adds new object information toupdate the template. Compared with the virtual template based template updating, thetemplate updating using Graph Cut to extract the foreground object information utilizesobject’s depth cues and appearance cues to update the template according to the objectfeature pool.Through the research work presented in this thesis, the robustness of the trackingperformance is improved which can be manifested in the Mean Shift algorithmicconvergence, running speed and accuracy.