一种基于多特征模型的Mean Shift跟踪算法  

A Mean Shift Tracking Algorithm Based on Multi-feature Model

作　　者：孙荣春[1] 孙俊喜[1] 宋雪[2] 

机构地区：[1]长春理工大学电子信息工程学院,长春130021 [2]吉林省机电研究设计院,长春130022

出　　处：《半导体光电》2011年第5期733-736,共4页Semiconductor Optoelectronics

摘　　要：计算机视觉领域中,目标跟踪技术有着广泛的实用价值。在复杂背景下要准确和稳定地实现目标跟踪,势必需要多信息融合技术。文章针对传统的基于颜色概率模型的Mean Shift算法忽略了目标空间信息这一不足,提出了联合特征的Mean Shift算法。文中将跟踪窗内子图像进行多级小波分解,用多级小波子带系数的统计特性构成纹理特征向量,再加权融合颜色概率直方图特征向量作为最终匹配特征向量。实验结果表明,在复杂背景下,该方法比传统基于颜色概率直方图模型的Mean Shift算法在准确性和鲁棒性上均有所提高。In the field of computer vision,the technology of object tracking has wide applications.To track an object accurately and robustly under complex background,multi-information fusion technology is required.The traditional Mean Shift algorithm based on the color probabilistic model ignores the spatial information,which affects the performance of the object tracking.An improved Mean Shift algorithm based on the wavelet transform was proposed,in which the sub-image of the tracking window was decomposed by the multi-level wavelet transform,and the statistical properties of the multi-level wavelet coefficients were used to constitute a texture feature vector.Then it was weighted and fused with the color histogram feature vector as the final matching feature vector.The experimental results show that this method obtains better accuracy and robustness compared with the traditional Mean Shift based on color probabilistic model.

关 键 词：小波变换 目标跟踪 Mean SHIFT 多特征 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]