Abstract:

When the shape,direction or color of the target changes,the traditional mean shift tracking algorithm often drifts and even fails because of its kernel window with fixed size and direction as well as its constant target model.In order to solve this problem,this paper proposes a new method that can adaptively adjust the size and direction of the kernel window,and presents a mechanism to update the target model.In the investigation,first,based on the ellipse obtained by fitting the convex hull of the target,the Kalman filtering model is used to calculate the optimal estimation of the target scale and orientation.Then,according to the estimation of the target scale,the size and direction of the kernel window are adjusted,and the distribution of the kernel weight is corrected.Finally,according to the shape and color information,a target model updating mechanism adaptive to the target change is presented.The proposed algorithm is applied to the tracking of video sequences of humans and non-motor vehicles under different scenes.The results show that the algorithm can accurately and stably track the target with obvious changes of scale,orientation and color.

Key words: target tracking, mean shift, kernel window, Kalman filtering, target model, updating