Research On Moving Object Tracking Algorithm Based On Mean Shift

With the rapid development of computer technology,image processing and computing capabilities have been continuously improved,and computer vision has also achieved rapid development.The tracking of moving objects involves many fields such as image processing,pattern recognition,and artificial intelligence.It is a cross-cutting discipline.Mean Shift algorithm is one of the basic algorithms for object tracking.Mean Shift algorithm needs to manually select the tracking target in the initial frame,and establish a corresponding histogram,reiteratively searching for subsequent video frames,When the selected similarity coefficient satisfies the set threshold or the number of iterations reaches the upper limit,the search is stopped and the position of the tracking target is determined.The point of Mean Shift algorithm is that it can converge to the approximate position of the target with a small number of iterations,and the calculation amount is small,and it has certain robustness to the target deformation.But its shortcomings and deficiencies are also obvious:(1)A single color feature cannot contain target spatial information and a single measure of similarity.(2)Tracking window scale fixed.(3)It is easy to trace failures in case of occlusion.So the scope of application is limited.In this paper,the following improvements have been made to the deficiencies of the Mean Shift algorithm:Firstly,an improved similarity measure function BJSD(Bhattacharyya coefficient and Jensen-Shannon Divergence)is proposed for the.The core idea is to introduce a spatial histogram and then calculate the similarity of the pixel's color information(that is,the first-order spatial information)and the second-order spatial information.The similarity between the two spatial distributions is calculated using the Jensen-Shannon Divergence,and the similarity of the color features is measured using the Bahrain distance.Finally,the two similarity measure methods are combined and applied to the Mean Shift algorithm.Experiments show that: compared to the traditional Mean Shift algorithm,The improved Mean Shift algorithm contains spatial information about the color distribution of the target area.Improves the shortcomings of the missing spatial position information of the pixel in the traditional Mean Shift algorithm.Then we improved the shortcomings of the Mean Shift algorithm tracking window,a coarse-to-fine two-step estimation method is proposed for the problem of fixed scale,which avoids the effects of scale mutation and small-scale loitering on the accuracy of tracking.first of all,for the update of the target's location,you can use improved spatial histogram similarity measures,select the most similar candidate template for updating;Secondly,in order to solve the problem of scale estimation mutation and the problem of "small-scale loitering"(the scale will maximize the similarity within an interval),the scale estimation of the constraint item is introduced,ie,the penalty item is added;Finally,a reverse-scale consistency check is performed.That is,the previous frame is backtracked to obtain the backward prediction scale.When the reverse consistency check is not satisfied,the estimated value is not reliable,and the previous frame's scale estimate is taken.Experiments show that the improved algorithm can effectively improve the target size adaptive problem,and the algorithm is robust.Finally,for the occlusion problem in target tracking,a combination of extended eurman filtering and Mean Shift algorithm is proposed.The target position is predicted based on the ability of the extended Kalman filter to estimate the motion system.The target position is predicted based on the ability of the extended Kalman filter to estimate the motion system.The target position is predicted based on the ability of the extended Kalman filter to estimate the motion system.