计算机科学与探索 ›› 2017, Vol. 11 ›› Issue (2): 294-302.DOI: 10.3778/j.issn.1673-9418.1511078

• 人工智能与模式识别 • 上一篇    下一篇

多特征的双人交互动作识别算法研究

黄菲菲1,曹江涛1,姬晓飞2+,王佩瑶1   

  1. 1. 辽宁石油化工大学 信息与控制工程学院,辽宁 抚顺 113001
    2. 沈阳航空航天大学 自动化学院,沈阳 110136
  • 出版日期:2017-02-01 发布日期:2017-02-10

Research on Human Interaction Recognition Algorithm Based on Mixed Features

HUANG Feifei1, CAO Jiangtao1, JI Xiaofei2+, WANG Peiyao1   

  1. 1. School of Information and Control Engineering, Liaoning Shihua University, Fushun, Liaoning 113001, China
    2. School of Automation, Shenyang Aerospace University, Shenyang 110136, China
  • Online:2017-02-01 Published:2017-02-10

摘要: 运动特征的选择直接影响基于整体的双人交互动作识别算法的识别效果。单一的特征因其适应范围不同,受到人体的外观、环境、摄像机设置等因素的影响,识别效果往往不太理想。在研究双人交互动作的表征与识别的基础上,充分考虑不同特征的优缺点,提出了一种结合局部的光流特征、局部的剪影特征以及HOG(histogram of oriented gradient)特征的混合特征,使用帧帧最近邻分类器获得3个特征的识别概率,最终通过加权融合3个特征的识别概率实现交互行为的识别。实验结果表明,对于UT-interaction数据库,该算法得到了较为理想的识别结果,混合特征可将识别率提高到91.7%。

关键词: 动作识别, 光流特征, 剪影特征, HOG特征

Abstract: The choice of motion features affects the result of human interaction recognition algorithm directly. Because of different adaptive scopes, many factors often influence the single features, such as the appearance of human body, environment and camera setting. So it can’t achieve satisfactory accuracy of action recognition. On the basis of studying the representation and recognition of human interaction action, and giving full consideration to the advan-tages and disadvantages of different features, this paper proposes a mixed feature which combines local optical flow feature, local silhouette feature and HOG (histogram of oriented gradient) feature. The nearest neighbor classifier is used to obtain the recognition probability of three features. Finally, the recognition result is achieved by weighted fusing those recognition probabilities. The experimental results demonstrate that this algorithm achieve better recognition results in the UT-interaction database, and compared with the single features, the mixed feature can improve the recognition rate to 91.7%.

Key words: action recognition, optical flow feature, silhouette feature, HOG feature