A video feature for action recognition was proposed. By observing 2D videos of human movement, different movement behaviors show different telescopic changes in human body and outline to some degree. Body outside silhouette and inner silhouette were termed as current frame poses, and variable poses movement. The pose-change-sequence frequencies and time mean squared errors were gathered to construct the eigenvectors. Several classification methods such as cross validating,features selecting and linear discriminant analysis were conducted on the collected data. The experimental results show that the eigenvectors have good linear separability, are un-sensitive, contain appropriate distinction information, and have higher recognition precision.%提出了一种行为识别的视频特征.观察人运动的2D视频,不同的运动行为在一定程度上表现为人体内外轮廓不同部位的伸缩变化.以每一帧人运动前景的内、外轮廓凸凹形状来表征当前帧的姿态,以姿态的变化来表征运动.采集姿态变化序列频率与时间平均方差构成的特征向量,利用多种分类方法对采集数据进行交叉检验、特征选择分析和线性判别分析.实验表明特征向量线性可分性好,对人是否背负物品不敏感,包含了恰当的行为区分信息,行为识别精度较高.
展开▼
