期刊文献+

融合视点机制与姿态估计的行人再识别方法 被引量:5

Person Re-identification Fusing Viewpoint Mechanism and Pose Estimation
下载PDF
导出
摘要 行人再识别是视频监控中一项极具挑战性的任务。图像中的遮挡、光照、姿态、视角等因素,会对行人再识别的准确率造成极大影响。为了提高行人再识别的准确率,提出一种融合视点机制与姿态估计的行人再识别方法。首先,采用姿态估计算法Openpose定位行人关节点;然后,对行人图像进行视图判别以获得视点信息,并根据视点信息与行人关节点位置进行局部区域推荐,生成行人局部图像;接着,将全局图像与局部图像同时输入CNN提取特征;最后,采用特征融合网络将全局与局部的特征融合,以获取更具鲁棒性的特征表示。实验结果表明:提出的方法具有更高的行人再识别准确率,其在CHUK03数据集上的rank 1达到了71.3%,在Market1501和DukeMTMC-reID数据集上的mAP分别达到了63.2%与60.5%。因此,所提方法能够很好地应对行人姿态变化和视角变化等问题。 Person re-identification is a very challenging task in video surveillance.Person have significant changes in appearance due to occlusion and differences in illumination,posture and perspective,which will ultimately have a great impact on the accuracy of person re-identification.To overcome these difficulties,this paper proposes a method for person re-identification based on viewpoint mechanism and pose estimation.First,the pose estimation algorithm Openpose is used to locate the joint points of person.Then,view discrimination is performed on the image to obtain viewpoint information.Local regions based on viewpoint information and joint point locations is proposed to generate a partial image.Next,the global image and the partial image are input into the CNN simultaneously to extract features.Finally,in order to obtain a more robust feature representation,the feature fusion network is used to fuse the global and local features.Experimental results show that the proposed method has higher person re-identification accuracy.On CHUK03 dataset,the value of rank1 reaches 71.3%,and on Market1501 dataset and DukeMTMC-reID dataset,the mAP reaches 63.2%and 60.5%,respectively.Therefore,the proposed methokd can well cope with person attitude changes,pose changes and other issues.
作者 裴嘉震 徐曾春 胡平 PEI Jia-zhen;XU Zeng-chun;HU Ping(College of Computer Science and Technology,Nanjing Tech University,Nanjing 211816,China)
出处 《计算机科学》 CSCD 北大核心 2020年第6期164-169,共6页 Computer Science
基金 国家自然科学基金(61672279) 国家重点研发计划(2017YFC0805605) 江苏省重点研发计划(BE2017617)。
关键词 相机视点 姿态估计 深度学习 行人再识别 特征融合 Camera viewpoint Pose estimation Deep learning Person re-identification Feature fusion
  • 相关文献

参考文献1

二级参考文献10

  • 1Felzenszwalb P F,Huttenlocher D P.Pictorial structures for object recognition[J].International Journal of Computer Vision,2005,66 (1):55-79.
  • 2Andriluka M,Schiele S R B.Pictorial structures revisited:people detection and articulated pose estimation[C] // Irfan Essa.Proceedings of IEEE Conference on Computer Vision and Pattern Recognition,2009.Florida:IEEE Computer Society,2009:1014-1021.
  • 3RamananD.Learning to parse images of articulated bodies[C] //Neural Information Processing Systems,2007.Cambridge:MIT Press,2007:1129-1136.
  • 4Ferrari V,Marin-Jimenez M,ZissermanA.Progressive search space reduction for human pose estimation[C] //Computer Vision and Pattern Recognition,2008.Anchorage:IEEE Computer Society,2008:1-8.
  • 5Ferrari V,Marin-Jimenez M,Zisserman A.Pose search:Retrieving People Using Their Pose[C] //Computer Vision and Pattern Recognition,2009.Florida:IEEE Conference Publication,2009:1-8.
  • 6Eichner M,Marin-Jimenez M.2D Articulated Human Pose Estimation and Retrieval in (Almost)Uncon-strained Still Images[J].Int J Comput Vis,2012,99:190-214.
  • 7Felzenszwalb P,McAllester D,Ramanan D.Object detection with discriminatively trained part-based models[J].IEEE Pattern Analysis and Machine Intelligence,2009,32(9):1627-1645.
  • 8Wang Hao,Meng Fan-hui,Fang Bao-fu.Iterative Human Pose Estimation Based on A New Part Appearance Model[J].Applied Mathematics & Information Sciences,2014,8(1L):311-317.
  • 9贾慧星,章毓晋.基于梯度方向直方图特征的多核跟踪[J].自动化学报,2009,35(10):1283-1289. 被引量:24
  • 10费兰英,范宁宁.基于人体图模型和背景减除的姿势估计[J].苏州大学学报(工科版),2012,32(3):8-14. 被引量:2

共引文献3

同被引文献15

引证文献5

二级引证文献4

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部