计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

基于改进密集轨迹的人体行为识别算法

程海粟 1,李庆武 1,2,仇春春 1,郭晶晶 1   

  1. (1.河海大学 物联网工程学院,江苏 常州 213022; 2.常州市传感网与环境感知重点实验室,江苏 常州 213022)
  • 收稿日期:2015-07-22 出版日期:2016-08-15 发布日期:2016-08-15
  • 作者简介:程海粟(1991-),男,硕士研究生,主研方向为机器学习、模式识别;李庆武(通讯作者),教授、博士;仇春春、郭晶晶,硕士研究生。
  • 基金项目:
    国家自然科学基金资助项目(41306089);江苏省产学研前瞻性研究基金资助项目(BY2014041);常州市科技支撑计划(社会发展)基金资助项目(CE20145038)。

Human Action Recognition Algorithm Based on Improved Dense Trajectories

CHENG Haisu  1,LI Qingwu  1,2,QIU Chunchun  1,GUO Jingjing  1   

  1. (1.College of Internet of Things Engineering,Hohai University,Changzhou,Jiangsu 213022,China; 2.Changzhou Key Laboratory of Sensor Networks and Environmental Sensing,Changzhou,Jiangsu 213022,China)
  • Received:2015-07-22 Online:2016-08-15 Published:2016-08-15

摘要: 针对基于轨迹技术的人体行为识别算法中轨迹提纯与特征表达有效性不足等问题,提出一种改进的人体行为识别算法。对视频进行运动显著性检测并提取传统的密集轨迹,通过分析当前帧和相邻帧的密集轨迹运动显著性值进行提纯。密集轨迹特征包括轨迹的位移向量和轨迹包络中每个时空块内的梯度方向直方图、光流直方图和运动边界直方图描述符。为更好地进行特征表达,根据运动显著性值分布优化词袋模型以获得更精确的视觉词典。在KTH和UCF sports数据集上的实验结果表明,该算法能够有效地提高识别率。

关键词: 行为识别, 运动显著性, 密集轨迹特征, 轨迹提纯, 词袋模型

Abstract: Aiming at the problem that recent human action recognition algorithms based on trajectories are defective is in trajectory refinement and feature representation,this paper proposes a novel human action recognition algorithm using improved dense trajectories.It detects the motion saliency and extracts the traditional dense trajectories in the videos,and refines the dense trajectories via the analysis of the motion saliency of current frame and adjacent frames.The dense trajectories features are formed by the sequence of displacement vectors in trajectories together with Histogram of Oriented Gradient(HOG),Histograms Optical Flow(HOF)and Motion Boundary Histograms(MBH)descriptors computed in each spatio-temporal patch.In order to represent the videos better,the model of Bag of Words(BOG) model is optimized according to the saliency value distribution to get the more accurate visual vocabulary.Experimental results show that the proposed algorithm can improve the human action recognition accuracy effectively on KTH and UCF sports action datasets.

Key words: action recognition, motion saliency, dense trajectory features, trajectory refinement, bag-of-words model

中图分类号: