核窗口尺寸和目标模型自适应的均值漂移跟踪

Mean Shift Tracking with Adaptive Kernel Window Size and Target Model

在目标大小、方向和颜色发生变化时,传统的均值漂移算法会因为核窗口大小和方向不能动态改变、目标模型不能及时更新而导致目标跟踪偏移甚至丢失.为此,文中提出了一种新的核窗口大小和方向可自适应调整的均值漂移跟踪算法,并构建了目标模型更新机制.首先利用计算得到的目标凸包拟合椭圆并结合卡尔曼滤波模型得到目标大小和方向的最优估计;然后利用目标大小和方向的估计值调整算法核窗口的大小和方向,修正核权重分布;最后联合目标形状和颜色信息构建一种目标更新机制,及时更新目标模型以适应目标的变化.不同场景下人体、非机动车等非刚体目标的视频序列实验结果表明,文中方法可以对大小、方向和颜色变化明显的目标进行准确、稳定的跟踪.

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.