1人体行为描述-中国图象图形学报.doc

中图法分类号:TP391.4 文献标识码:A 文章编号:1006-8961(2014) – 140537 论文引用格式：李俊峰, 张飞燕. 基于局部时空特征方向加权的人体行为识别[J].中国图象图形学报，2014 基于局部时空特征方向加权的人体行为识别 李俊峰，张飞燕 浙江理工大学 自动化研究所，浙江 杭州 310012 摘 要：目的：对人体行为的描述是行为识别中的关键问题，为了能够充分利用训练数据从而保证特征对行为的高描述性，本文提出了基于局部时空特征方向加权的人体行为识别方法。方法：首先，将局部时空特征的亮度梯度特征分解为三个方向（、、）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三方向特征描述子集合的标准视觉词汇码本，并利用训练视频得到每个行为的标准三方向词汇分布；进而，根据不同行为三方向特征描述子集合的标准视觉词汇码本，分别计算测试视频相应的三方向的词汇分布，并利用与各行为标准三方向词汇分布的加权相似性度量进行行为识别；结果：最后，本文在Weizmann数据库和KTH数据库中进行实验，Weizmann数据库中的平均识别率高达96.04%，KTH数据库中的平均识别率也高达96.93%。结论：与其它行为识别方法相比可以明显提高行为平均识别率。 关键词：行为识别；局部时空特征；视觉词汇表；方向加权 Recognition of human behavior based on the directional weighting local space-time features Li Junfeng, Zhang Feiyan Institute of automation ,Zhejiang Sci-Tech University, Hangzhou 310012,China Abstract：a new method of human activity recognition is put forward in this paper. Method: Firstly, the brightness gradient was decomposed into three directions (、、) which could describe the behavior from different directions respectively. Secondly, their standard visual vocabulary codebooks of three directions for different behavior could be obtained by constructing visual vocabulary directly. Moreover, based on the standard visual vocabulary codebooks of the three directions for each behavior, the corresponding vocabulary distributions of the test video were calculated separately. And the behavior of the test video might be recognized by using the weighted similarity measure between each behavior’s standards vocabulary distribution and vocabulary distribution of the test video. Result: Finally, the performance was investigated in the KTH action dataset and Weizmann action dataset. And we had obtained the average recognition rate of 96.04% accuracy in Weizmann action dataset and 96.93% accuracy in KTH action dataset. Conclusion: Our method could generate a more comprehensive and effective representation for action videos. And, it can reduce the clustering time by producing codebooks of each direction. The experi