欢迎您！东篱公司
退出
-
申报数据库
-
立项数据库
-
成果数据库
-
项目获奖数据库
中文摘要：
在视频跟踪任务中，以上一帧跟踪到的目标位置为基础，在当前帧内相应位置周围生成若干候选区域样本进行分类，并从中获取待跟踪目标在当前帧中的位置和更新分类器，这是基于判别式方法的基本跟踪流程．对于每帧产生的大量未标记类别的候选区域样本，现有的基于子空问学习的跟踪方法大多忽略了这些样本内在的几何结构，而是直接向子空间投影，并在子空间内进行二分类，区分出其中的正类样本（前景）和负类样本（背景）．在半监督判别分析方法的基础上，提出一种基于增量半监督判别分析的跟踪方法框架．首先，使用区域协方差特征描述子提取图像中不同区域的大量图像特征；然后，为保持这些特征间的几何结构，将它们映射至欧氏空间内进行处理；再将原始半监督判别分析方法扩展到增量形式，给出类内散度矩阵、类间散度矩阵和正则项的增量更新方法，并由此给出目标跟踪的流程框架；通过实验显示，该方法对于目标跟踪问题具有良好的实时性和鲁棒性．
英文摘要：
In the task of video tracking, based on the location of object in last frame, generating several candidate regions as samples for classification and extracting the location of object in current frame with maximum a posterior （MAP）, are the framework of discriminant tracking（one of two common tracking framework, the other one is generative tracking）. For those unlabeled candidate samples generated in every new coming frames, existing subspace learning based tracking methods often ignore the inherent structures of those samples, map them to the subspace directly, and distinct positive samples（objects） from negative ones（backgrounds） in projecting subspace.Based on semi-supervised discriminant analysis （SDA）, an tracking framework based on incremental extended version of SDA is proposed in this paper. Firstly, considering the efficiency of tracking, we use a fast feature descriptor for feature extraction, which is called region covariance matrix（RCM）. With RCM, 7-dimension vector can be obtained for every pixel in interesting regionsi secondly, in order to preserve the geometric structures among region covariance matrix features during tracking, we map them from symmetric positive definite manifold space to Euclidean space; then, original SDA is extended to the incremental version, where within-class, between-class scatter matrixs and regular terms are ineremently extended, and a framework for obiect tracking is proposed; finally, experimental results show the efficiency and robustness of our method.
同期刊论文项目
同项目期刊论文
An optimal symmetrical null space criterion of Fisher discriminant for feature extraction and recogn
期刊信息
位置：