基于均值移动与贪婪算法的多人脸跟踪

摘要

提出一个基于均值移动(Mean Shift)和贪婪算法的多人脸跟踪器.首先建立多个均值移动目标跟踪器以进行多人脸跟踪.结合卡尔曼滤波逐个检测目标并从视频帧中清除已跟踪到的人脸,以解决当多个目标相邻或相互遮挡时相应的跟踪窗口会收敛于最大目标、导致其他目标丢失的难题.引入辅助窗口并根据其纹理信息确定粘连目标的对应.实验结果表明,该多人脸跟踪算法可实现稳健的实时多人脸跟踪.%A novel multiple face tracking algorithm is proposed in this paper. Multiple Mean Shift trackers are first built to enable multiple face tracking. To overcome the weakness of Mean Shift tracking, which is prone to converge to the local maximum target if tracked objects are adjacent or partially occluded, a greedy tracking method is used to pursue the targets one by one, during which a Kalman filter is first employed to locate the initial position, and then the tracked object is removed from the scene to guarantee no other Mean Shift tracker iterates the same target. An accessory window featured with local texture distribution is introduced to correspond to candidate widows and targets. Experimental results have indicated the proposed algorithm can track multiple faces robustly in real time.