Research On Algorithm Of Video Target Tracking Based On Mean Shift

Video target tracking is an important research orientation of computer vision area. It has important research value in science and engineering, production and life, which are widely used in intelligent human-computer interaction, medical diagnostics, intelligent robot, video surveillance and other areas. The Mean Shift algorithm is a fast and efficient technique for climbing density gradients to find the steady point, which has become one of research hot spots in video target tracking algorithm.In this dissertation, the research is focus on algorithm of video object tracking based on Mean Shift. Tracking objects is the interest target block in the picture of video. Three improved tracking algorithm have been proposed with summarizing and analysis on already developed algorithms. The three algorithms are named respectively as:CAM Shift object tracking algorithm based on color and edge feature, adaptive kernel bandwidth Mean Shift target tracking algorithm, multi-feature adaptive kernel bandwidth Mean Shift target tracking algorithm. These visual tracking algorithms can accurately estimate not only the position of target, but also the scale and orientation of target.The main research contents and contributions of this dissertation are summarized as follows:1. The continuously adaptive mean shift (CAM Shift) algorithm is vulnerable to similar color interference and light changes affect, to address this problem, the new algorithm developed here based on multidimensional color space and edge feature. Firstly, weight the color space histogram with decreasing kernel profile function to compute the probability of the color feature value belonging to the target model. Then, detect target edge feature with Sobel operator, the probability distribution image of candidate target is determined by giving color and edge different weights. Finally, find the target with CAM Shift algorithm. The experiments show that the proposed algorithm can track multi-color target and gets better performance in similar color background.2. The kernel function bandwidth of the traditional Mean Shift target tracking algorithm can’t adapt to the changes of scale and orientation of the target, to address this problem, adaptive kernel bandwidth Mean Shift tracking algorithm is proposed. Firstly, establish the target model with the color space histogram, thus, the probability density distribution image of target is established in the current image by mapping target model, then, with calculating the zeroth moment of the probability density, the target window width and kernel-bandwidth are adaptive adjusted. Target the tracking object with the ellipse, the ellipse parameters obtained by moment’s operation. The algorithm developed here can not only adapt to target scale changes, but also to estimate the target orientation, because the dynamically changing probability distribution blob can be calculated. Tracking experiments verify the effectiveness of this algorithm. It has better advantages than traditional Mean Shift tracking algorithm in change of scale and direction, more robust than CAM Shift algorithm in similar color interference.3. Considering the simple target feature can’t accurately describe the target, multi-feature adaptive scale Mean Shift tracking algorithm is proposed. The algorithm use color and texture information to create the target feature model. Firstly, create the target probability density distribution using target feature model in optimal target location, and then, calculate the scale of density distribution, thus, the next frame target scale is adaptive adjusted according to the value. Finally, target the object with ellipse, the ellipse parameters is obtained by calculating first moment and second moment of probability density in optimal location. The algorithm successfully achieves the position, scale and orientation of target.