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

• 图形图像处理 • 上一篇    下一篇

静止背景下的人体行为识别方法

高晨兰,朱嘉钢   

  1. (江南大学 物联网工程学院,江苏 无锡 214122)
  • 收稿日期:2016-10-09 出版日期:2017-10-15 发布日期:2017-10-15
  • 作者简介:高晨兰(1992—),女,硕士研究生,主研方向为图形图像处理、人工智能、模式识别;朱嘉钢,副教授、博士。
  • 基金资助:
    江苏省产学研项目(BY2013015-40)。

Human Actions Recognition Method Under Stationary Background

GAO Chenlan,ZHU Jiagang   

  1. (School of Internet of Things Engineering,Jiangnan University,Wuxi,Jiangsu 214122,China)
  • Received:2016-10-09 Online:2017-10-15 Published:2017-10-15

摘要:

在实际工程中多数监控摄像头是固定的,为了使计算能力有限的智能摄像头实时地进行人体行为识别,提出一种将无迭代双边二维主成分分析方法(NIB2DPCA)与高斯混合模型(GMM)相结合的行为识别方法。提取视频帧序列中运动前景的稠密光流,绘制运动矢量时空(MVFI)模板,利用NIB2DPCA对MVFI模板作特征抽取,通过GMM对特征数据建模从而实现行为分类。测试结果表明,与轨迹云比较法相比,该方法对视频中的行为信息进行了有效地压缩,使得计算耗时缩短了90%以上,同时保持了较高的识别率。

关键词: 人体行为识别, 运动特征, 运动矢量, 特征抽取, 高斯混合模型

Abstract:

Aiming at the situation that most surveillance cameras are fixed in practical engineering applications,in order to enable the intelligent cameras with limited computing ability to identify human actions in real time,an action recognition method combining Non-iteration-bilateral-projection Based Two Dimensional Principal Component Analysis(NIB2DPCA) and Gaussian Mixture Model(GMM) is proposed.The dense optical flow of the motion foreground in the video frame sequence is extracted to draw the Motion Vector Flow Instance(MVFI) template.NIB2DPCA is employed for feature extraction from MVFI templates.The feature data is modeled by GMM to construct the classifier for recognizing human actions.Test results show that compared with the trajectory cloud comparison method,the proposed method which effectively compresses the action information in the video reduces the calculation time by more than 90% while maintaining a high recognition rate at the same time.

Key words: human actions recognition, motion feature, motion vector, feature extraction, Gaussian Mixture Model(GMM)

中图分类号: