基于光流速度分量加权的人体行为识别

Human Behavior Recognition Based on Weighted Optical Flow Velocity Component

为了提高光流特征对人体行为的描述性,提出一种新的人体行为识别方法。首先,将提取的光流特征分解为u和v两个速度分量分别来描述行为,通过直接构造视觉词汇表分别得到不同行为两个速度分量的标准视觉词汇码本,并利用训练视频得到每个行为的不同分量的标准词汇分布;进而根据不同行为两个分量的标准视觉词汇码本,分别计算测试视频相应的速度分量的词汇分布,并利用与各行为两个速度分量的标准词汇分布的加权相似性度量进行行为识别;最后在KTH数据库和Weizmann数据库中进行实验。实验结果表明,与其它行为识别方法相比笔者提出的方法可以明显提高行为平均识别率。

In order to improve the descriptiveness of optical flow characteristic on human behavior, this paper proposes a kind of new human behavior recognition method. Firstly, the extracted optical flow characteristic is decomposed as 2 velocity components （u and v） to describe the behavior respectively, and then the standard visual vocabulary code book of 2 velocity components can be obtained via direct construction visual vocabulary table, and the testing video is utilized to obtain the distribution of standard vocabulary of different components of each behavior then, according to standard visual vocabulary code book of 2 components of different behavior, it is able to respectively calculate the vocabulary distribution of corresponding velocity component of testing video, and utilize the weighted similarity measurement of standard vocabulary distribution of 2 velocity components of each behavior to carry out behavior recognition; finally, the test is made in KTH database and Weizmann database; through comparison between experimental result and other behavior recognition methods, it can be found that the average recognition rate of behavior is obviously improved.