增强目标模型鲁棒性的Mean-shift算法

Mean-shift Tracking with Enhanced Robustness of Target Model

针对传统Mean-shift算法中颜色核函数直方图对目标特征描述较弱的缺点,提出了一种联合目标特征点的二维结构信息和颜色信息的Mean-shift改进算法.改进算法细化了Harris检测算子的角点响应阈值,提取出更多的目标特征点计算其方向分布,并以方向与部分颜色特征的直方图构建目标模型,该模型能显著区分目标与背景.实验对不同算法进行了仿真及性能比较,结果表明:提出的改进算法在一定的复杂场景中提高了跟踪精度,且具有较好的鲁棒性.

The traditional Mean-shift cannot represent accurately the color distribution of the object. To tackle this problem, a target motion-tracking algorithm is presented based on two-dimensional structural information of the target feature points and color histogram. The improved algorithm refines the Harris operator threshold to seek more feature points, with which their orientation distribution can also be sought. The target model built by combining the orientation histogram and the partial color histogram is found to be able to separate target from background to a satisfying extent. The application results from various video sequences suggest good localization precision in object tracking, and algorithmic robustness in complex scenes.