视频中人体行为的慢特征提取算法

PDF下载HTML

Title:

Slow feature extraction algorithm of human actions in video

作者:

陈婷婷^1,2, 阮秋琦¹, 安高云¹

1. 北京交通大学 信息科学研究所, 北京 100044;
2. 北京交通大学 现代信息科学和网络技术北京市重点实验室, 北京 100044

Author(s):

CHEN Tingting^1,2, RUAN Qiuqi¹, AN Gaoyun¹

1. Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China;
2. Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing Jiaotong University, Beijing 100044, China

关键词:

人体行为; 训练立方体; 慢特征函数; 慢特征; 帧间差分法

Keywords:

human action; training cuboids; slow feature function; slow feature; frame difference

分类号:

TP391

DOI:

10.3969/j.issn.1673-4785.201407002

文献标志码:

A

摘要:

从复杂的人体行为中提取出重要的有区分力的特征是进行人体行为分析的关键.目前经典的特征分析方法大多是线性的特征分析技术,对于非线性处理会导致错误的结果,为此,提出了一种慢特征提取方法.首先,利用帧间差分法获取帧差图像序列,对选定的初始帧进行特征点检测;然后,利用光流法对特征点进行跟踪,收集训练立方体;最后,利用收集的训练立方体进行慢特征函数的机器学习,提取出慢特征并进行特征表示.实验中提取每种行为的慢特征进行对比,结果显示提取的慢特征随时间变化非常缓慢,并且在不同行为之间具有很强的区分力,表明该方法能够有效提取出人体行为的慢特征.

Abstract:

Extracting important and distinguishable features from complex human actions is the key for human actions analysis. In recent years, classical feature analysis methods are mostly linear feature analysis technologies, which result in error results for non-linear processing. This paper proposes a method of extracting slow features. First, the image sequence of frame difference was obtained by the difference between the consecutive frames and some feature points of selected beginning frame were detected. Next, the feature points were tracked by optical flow method and the training cuboids were collected. Finally, the slow feature functions were learned with the collected training cuboids, then the slow features could be extracted and represented. In the experiment, slow features of each action were extracted and compared with each other. The results show that the extracted slow features vary slowly with time and action interclass has good discrimination, which suggests that this method can extract slow features from human actions effectively.