中文摘要：

提出利用梯度和光流的统计特征进行人体行为识别的新方法.首先,通过数理统计分析得到不同行为的梯度和光流的直方图分布均符合非对称广义高斯分布（AGGD）;然后,分别提取梯度和光流的AGGD模型的参数,并把这些参数作为描述人体行为的统计特征;最后,通过计算训练集行为视频与测试集行为视频的统计特征间的马氏距离进行人体行为识别。在KTH数据库和Weizmann数据库上分别进行了实验仿真,两个数据库上的平均识别率分别高达95.20%和93.16%,与其它行为识别方法相比可以明显提高行为平均识别率。

英文摘要：

Through the analysis of histogram distribution of the local spatio-temporal features （gradient and optical flow） for different behavior videos, it is found that the statistics characteristics of gradient and optical flow for different behavior videos are obviously different respectively. In order to ensure the high descriptive of features to the behavior, a new method of human activity recognition is put forward by using the statistics characteristics of gradient and optical flow in this paper. Firstly, it is found that the histogram distributions of gradient and optical flow for different behavior videos conform to the asym- metric generalized Gaussian distribution （AGGD） through the mathematical statistic analysis. Secondly, the parameters of AGGD model are extracted respectively and fused to describe different behavior as the statistical features. Moreover,human behavior is recognized through calculating the Mahalanobis distance between the test videors feature matrix and the train videosrl Finally,the performance is investigated in the KTH action dataset and Weizmann action dataset, and the average recognition rates are as high as 93. 16% and 95.20% for the two action datasets, respectively. The results show that this method can generate a more comprehensive and effective representation for action videos.