Research On Mean Shift Theory And Application On Image Processing And Object Tracking

The mean shift algorithm is an iteration algorithm, which has been widely used in the image processing and the computer vision, still has some theoretical proficiency. Therefore, the theoretical proven of the mean shift algorithm is perfected as well as its application in the image smoothing, image segmentation and object tracking is studied in the paper. The main finished tasks in this paper are as fellows:In view of the deficiency in the analysis of the density function of the mean shift algorithm, the relation of the concavity and convexity between the profile function k(x) and the kernel function k(||x-c||~2), the kernel k(||x - c||~2) and the density function f(x) are discussed. Besides, the iterative step and time of two motion rules of the mean shift algorithm are compared. Some theorems are analyzed and proposed, such as the convergence of the mean shift algorithm is proved using the Cauchy theorem; the one that the angle of the mean shift vectors in the same window is less than 90 degree,etc. some conditions are summarized while the condition of the mean-shift step is Newton step, quasi-newton step or other step; the local maximum value of density in the moving direction is discussed; and the necessary and sufficient condition is finally analyzed that the first derivative of the density function is equal to zero on the convergence point, etc.At last, the application of the mean shift algorithm in image smoothing, image segmentation and object tracking is tested; the image smoothing algorithm based on the cylindrical HSV space is studied in virtue of image smoothing problems; the optimal combination methods based on the multiple same scene images in the cylindrical HSV space is proved; the effect of the image smoothing is discussed with different color and space steps. Considering the image segmentation problems, the object segmentation algorithm based on the image edges of the current frame (the edge is gotten based on the mean shift algorithm) and the optimal difference template is proposed. With a view to object tracking problems, the two-window feature extracting algorithm, the object feature updated algorithm based on the pixel, and the linear combination method of the tracking windows are proposed separately.