Abstract: Mean shift algorithm of moving target tracking can not always adapt to the change of the target scale because it works with an invariable bandwidth kernel during the real tracking scenes. An adaptive moving target tracking algorithm based on fuzzy inference system is proposed in this paper. The location of target in the present frame is predicted by Kalman filter. The target scale is on-line adjusted adaptively by the fuzzy inference system where log likelihood ratio for object versus background and the similarity between candidate target and model is as its input. The target is tracked by Mean shift iterations. A model updating mechanism by the similarity which combines with the log likelihood ratio is designed to update the model of the tracking target adaptively. The algorithm is robust and tracks accurately. The experimental results justify that the algorithm can adapt to the change of the target scale and the background, and the tracking precision is superior to traditional Mean shift algorithm.

Key words: adaptive tracking, Mean shift algorithm, fuzzy inference system, Kalman filter

摘要： 运动目标在跟踪过程中往往伴随着尺度、形状的变化，Mean shift跟踪算法由于采用固定的核窗宽度进行运动目标跟踪，因而它本身不能适应这种变化。针对Mean shift算法存在的缺点，提出一种基于模糊推理的自适应Mean shift跟踪算法，该算法利用卡尔曼滤波算法对目标当前位置进行预测；设计模糊判定准则在线调整目标尺度值，利用Mean shift迭代运算逐步逼近目标完成跟踪；利用相似度和置信度系数设计模型更新准则，以实现模板的自适应更新。实验结果证明，该算法能够适应目标尺度和背景的变化，较普通的Mean shift跟踪算法不仅跟踪精度提高，而且鲁棒性更强。

关键词: 自适应跟踪, Mean shift, 模糊判决方法, 卡尔曼滤波