作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (8): 268-272,278. doi: 10.19678/j.issn.1000-3428.0047121

• 图形图像处理 • 上一篇    下一篇

基于LLC与GIST特征的静态人体行为分类

王恩德 1,2,刘巧英 1,2,3,李勇 1,2,3   

  1. 1.中国科学院沈阳自动化研究所,沈阳 110016; 2.中国科学院光电信息处理重点实验室,沈阳 110016; 3.东北大学 信息科学与工程学院,沈阳 110819
  • 收稿日期:2017-05-09 出版日期:2018-08-15 发布日期:2018-08-15
  • 作者简介:王恩德(1980—),男,研究员、博士,主研方向为图形图像处理、模式识别、机器人视觉;刘巧英,硕士研究生;李勇,博士研究生。
  • 基金资助:

    国家自然科学基金(61401455)。

Static Human Behavior Classification Based on LLC and GIST Features

WANG Ende 1,2,LIU Qiaoying 1,2,3,LI Yong 1,2,3   

  1. 1.Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China; 2.Key Laboratory of Optical Electrical Image Processing,Chinese Academy of Sciences,Shenyang 110016,China; 3.College of Information Science and Engineering,Northeastern University,Shenyang 110819,China
  • Received:2017-05-09 Online:2018-08-15 Published:2018-08-15

摘要:

针对静态图像人体行为识别问题,提出一种融合局部约束线性编码(LLC)和全局特征描述子的方法。该方法对图像进行密集采样,提取每个子区域的SIFT特征,利用LLC方法对提取的密集SIFT特征进行编码和池化。为了加入空间信息,采用空间金字塔的思想,获得具有空间位置 信息的LLC池化特征。将LLC池化特征串联通用搜索树(GIST)特征作为图像的最终描述,使用核函数为直方图交叉核函数的支持向量机进行分类。实验结果表明,与利用LLC、空间金字塔匹配特征和GIST特征进行识别的方法相比,该方法识别效果较好。

关键词: 行为识别, 全局特征描述子, 局部约束线性编码, 空间金字塔匹配, 最大池化

Abstract:

Aiming at the problem of static image human behavior recognition,a method of merging Local-constrained Linear Coding(LLC) and global feature descriptors is proposed.This method makes intensive sampling of the image,and extracts the Scale-Invariant Feature Transform(SIFT) features in each subregion.Then the LLC method is used to encode and pool the dense SIFT features.In order to add spatial information,spatial pyramid information is used to extract LLC pooling features with spatial location information.Generalized Search Trees(GIST) features are extracted,and the final features of the image are described using LLC pooling features tandem GIST features.The kernel function is the histogram cross kernel function Support Vector Machine (SVM) and it is used for classification.Experimental results show that compared with the methods using LLC,Spatial Pyramid Matching(SPM) features and GIST features,the proposed method has better recognition effect.

Key words: behavior recognition, global feature descriptor, Locality-constrained Linear Coding(LLC), Spatial Pyramid Matching(SPM), max pooling

中图分类号: