Mean-Shift跟踪算法中目标模型的自适应更新  

Mean-Shift Tracking with Adaptive Model Update Mechanism

作　　者：彭宁嵩[1] 杨杰[1] 周大可[1] 刘志[1] 

机构地区：[1]上海交通大学图像处理与模式识别研究所

出　　处：《数据采集与处理》2005年第2期125-129,共5页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(30170274l)资助项目;上海市科委人脸识别(03DZ14015)资助项目。

摘　　要：针对Mean-shift跟踪算法中的模型更新问题,提出利用目标历史模型和当前匹配位置处得到的观测模型,对目标核函数直方图进行Kalman滤波,从而对目标模型进行及时更新。在滤波过程中,通过分析滤波残差动态,调整滤波方程中的各种参数。Bhattacharyya系数被用作模型更新的准则。该系统能够有效地处理遮挡、光照变化等干扰,避免了模型的过更新。大量视频序列测试的结果表明,在场景遮挡、光照变化等因素的影响下,算法能够对目标外观以及尺度的变化进行稳健、实时和有效的跟踪。Considering the issue of model update within the Mean-shift framework, this paper proposes a model update method by using Kalman filters to estimate the object kernel-histogram from the previous and current object models. Filtering residuals are employed to adaptively tune all the parameters of Kalman filters. In addition, Bhattacharyya coefficient is used as a criterion for model update. Therefore, the tracker can avoid over-update. Experimental results show that the method keeps up with the object appearance and scale changes, and is robust under the influence of occlusion and illumination factors, etc.

关 键 词：目标跟踪 Mean—shift 自适应Kalman滤波 模型更新 BHATTACHARYYA系数 

分 类 号：TN911[电子电信—通信与信息系统]