摘要
为实现对关节式目标的稳定跟踪,提出了基于增量学习的关节式目标跟踪算法.该算法应用图割法对目标矩形窗进行前景与背景分割,得到前景图像;然后对前景图像进行快速傅里叶变换,得到傅里叶系数矩阵,进而得到振幅图像,并将振幅图像作为跟踪目标的描述;最后将多个目标描述进行奇异值分解和主元分析,实现对跟踪目标的低维子空间描述.文中在粒子滤波框架下实现了整个跟踪算法.实验结果表明,该算法具有较稳定的关节式目标跟踪效果.
In order to realize stable articulated object tracking,an algorithm based on incremental learning is proposed.In this algorithm,the graph-cut algorithm is used to obtain a foreground image by segmenting the rectangular object region,and a fast Fourier transform is conducted for the foreground image to obtain the Fourier coefficient matrix and to further acquire the amplitude image as the description of the tracking object.Then,the low-dimension subspace representation of the tracking object is obtained by the singular value decomposition and the principle component analysis of the amplitude image.Thus,the tracking algorithm is realized in the framework of particle filtering.Experimental results indicate that the proposed algorithm helps to achieve stable articulated object tracking.
出处
《华南理工大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2012年第3期88-93,共6页
Journal of South China University of Technology(Natural Science Edition)
基金
国家自然科学基金重点项目(60736024
60574004
61174053)
教育部科技创新工程重大项目(7080690)
关键词
目标跟踪
增量学习
子空间描述
快速傅立叶变换
奇异值分解
粒子滤波
object tracking
incremental learning
subspace representation
fast Fourier transforms
singular value decomposition
particle filtering
