摘要
针对视频跟踪过程中目标的形变、遮挡、旋转和背景干扰问题,提出一种融合超像素与动态图匹配的视频跟踪方法。首先,采用融合局部熵特征的简单线性迭代聚类(simple linear iterative clustering,SLIC)方法经聚类分析生成超像素集合,使生成的超像素边缘贴合度更好。其次,采用图像分割(graph cuts)方法生成候选目标超像素集合,并融合在线支持向量机学习算法(online SVM learning algorithm,LASVM)分类预测结果,使前景与背景分离的准确度更高。然后,充分利用目标的几何结构信息构建基于图模型的相似度矩阵,解决目标的形变和遮挡问题。理论分析与仿真结果表明:相比现有其他视频跟踪方法,新方法对跟踪过程中的遮挡和形变情况具有较强的鲁棒性,对一定程度的背景干扰和旋转问题跟踪效果良好。
Focusing on the problem of target deformation, occlusion, rotation and background interference, a video tracking method jointing superpixels and dynamic graph matching was proposed in this paper. Firstly, superpixels were generated by the simple linear iterative clustering analysis method integrating the local entropy feature, so that we can get superpixels that the edge fit better. Secondly, the candidate target superpixels was generated by graph cuts method, in which the LASVM classifier was combined with the graph guts method in order to make the separation of foreground and background more accurately. Thirdly, when the graph modle was constructed, we make full use of the geometric irttormation of the target to solve the problem of the occlusion and deformation effectively. Meanwhile, the constraints were introduced to reduce the dimension of the affinity matrix, so that the computational complexity was reduced. Theoretical analysis and simulation results show that our method has strong robustness and better tracking accuracy when deals with the occlusion and deformation, and the proposed method is good in dealing with the target rotation and a certain degree of background interference, compared with currently other video tracking methods.
作者
张君昌
周艳玲
万锦锦
Zhang Junchang Zhou Yanling Wan Jinjin(School of Electronics Information, Northwestern Polytechnical University, Xi＇an 710072, China Science and Technology on Electro-Optic Control Laboratory, Laoyang 471000, China)
出处
《西北工业大学学报》
EI
CAS
CSCD
北大核心
2017年第1期133-137,共5页
Journal of Northwestern Polytechnical University
基金
光电控制技术重点实验室和航空科学基金(2016515303)资助
关键词
目标追踪
信息融合
简单线性迭代聚类
超像素
图像分割
target tracking
information fusion
simple linear iterative clustering
superpixels
graph cuts
