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.
出处
《数据采集与处理》
CSCD
北大核心
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。
