正在加载图片...
第8卷第1期 智能系统学报 Vol.8 No.1 2013年2月 CAAI Transactions on Intelligent Systems Feh.2013 D0I:10.3969/j.issn.16734785.201209033 网络出版地址:htp:/nw.cmki.net/kcms/detail/23.1538.TP.20130125.1430.002.html 双视角下多特征信息融合的步态识别 李一波,李昆 (沈阳航空航天大学自动化学院,辽宁沈阳110136) 摘要:针对步态识别研究中单视角识别率低、多视角算法复杂等问题,开展了双视角下的步态识别研究.考察正面 视角人体的轮廓特征和侧面视角人体行走的动态特征,利用多视角步态信息互补性强的特点,分别从正面视角和侧 面视角获取步态序列,预处理得到单连通人体轮廓图形,然后对正面视角提取Procrustes均值形状,侧面视角计算动 作能量图(AEI)并经二维局部保留映射(2D-LPP)降维,最后将2个视角下的识别结果进行融合从而获得最终的识 别结果.在中科院自动化所的DatasetB数据库上进行了实验,获得了较高的识别率,达到了预期的识别效果. 关键词:步态识别;多特征信息融合;双视角;Procrustes均值形状;动作能量图;二维局部保留映射 中图分类号:TP391.41文献标志码:A文章编号:16734785(2013)01-0074-05 Gait recognition based on dual-view and multiple feature information fusion LI Yibo,LI Kun (College of Automation,Shenyang Aerospace University,Shenyang 110136,China) Abstract:In view of low recognition rate of single-view and complexity of multi-view algorithm,a research was con- ducted examining the gait recognition under dual-view.Current research on the contour characteristic of the human body in frontal view and the dynamic characteristics of human walking in side view was examined using the comple- mentary features of the gait information under multi-view.Also the gait sequences were obtained utilizing the two views respectively,and then preprocessed to obtain simply connected body silhouettes.Next,the Procrustes mean shape was extracted from the front view,and the active energy images(AEI)was calculated by side view.Howev- er,each of the AEI was projected to a low-dimensional feature subspace via two-dimensional local preserving pro- jections (2D-LPP).The final recognition result was obtained by fusing recognition results of two perspectives.The experiments in CASIA dataset(Dataset B)obtained a high recognition rate and achieved the expected effect of rec- ognition. Keywords:gait recognition;multiple feature information fusion;dual-view;Procrustes mean shape;active energy image;two-dimensional partial preserving projections 步态识别是生物特征识别领域一个具有广阔应 态识别研究的深入,涌现出许多算法.根据视角数目 用前景的研究方向,旨在根据人们走路的姿势进行 的不同,步态研究可分为单视角和多视角.单视角研 身份的识别,具有远距离、非侵犯性、易于感知、难于 究是指对单个视角下的视频序列进行特征提取和识 伪装等特点.鉴于步态识别的诸多优势及其在智 别.由于单视角下的步态序列存在遮挡、视角局限性 能监控、人机交互等领域的应用前景,广大科研工作 等影响因素,所能提供的步态信息有限,因此特征选 者积极投入相关的研究当中2].近年来,随着对步 取至关重要.Kusakunniran等3]提出了加权二值模 式(weighted binary patter,WBP)的方法;张浩等 收稿日期:2012-09-15.网络出版日期:201301-25 基金项目:国家自然科学基金资助项目(61103123). 提出了加权动态时间规整(dynamic time warping, 通信作者:李昆.E-mail:likun565@163.com. DTW)距离的自动步态识别算法;Bashir等[5]提出了
向下翻页>>
©2008-现在 cucdc.com 高等教育资讯网 版权所有