带宽自适应的Mean Shift目标跟踪算法

Bandwidth-Adaptive Mean-Shift Target Tracking Algorithm

针对传统均值漂移(Mean Shift)目标跟踪算法中核函数带宽缺乏良好自适应调整的缺点,提出了自适应调整核函数带宽的Mean Shift目标跟踪算法.该算法首先采用核函数计算目标颜色特征值的概率密度,在视频当前帧目标的最优位置区域由目标颜色特征概率投影生成目标概率密度分布图;然后根据概率密度零阶矩值调整下一帧跟踪窗口宽度,从而实现核函数带宽的自适应调整;最后通过矩运算计算椭圆参数,用椭圆锁定目标来实现复杂背景下目标的空间、尺度和方向定位.人脸跟踪实验结果表明:与传统MeanShift目标跟踪算法相比,文中算法可以实时地对目标进行缩放锁定,且能够估计目标姿态;与Cam Shift算法相比,文中算法抵抗相似颜色干扰的性能较好.

In this paper,a new mean-shift target tracking algorithm is proposed to improve the kernel function bandwidth-adaptive ability of the traditional one.First,the probability density of the eigenvalue of the target color is derived by employing the kernel function.Next,a distribution image of the target probability density is projected on the new optimal location of the target in the current video frame.Then,according to the zeroth-order moment of the probability density distribution,the width of the tracking window in the next frame is adjusted.Thus,the adaptive bandwidth of kernel function is achieved.Finally,the ellipse parameters derived by means of the moment ope-ration are adopted to lock the tracking target,thus achieving the target position in space,scale and direction in a complex background.Face-tracking experimental results show that,as compared with the conventional algorithm,the proposed one can achieve real-time scaling and locking of the target and estimate the target attitude,and that,it is superior to the Cam Shift algorithm in terms of resistance to the interference of similar color.