Abstract: Human abnormal activity recognition is studied based on approximate motion period of the human action.First of all，human contours are extracted by the use of background difference of Gaussian mixture model，and then the approximate period of human action is obtained by analyzing human shape’s deformation.After period analysis，human action sequence is decomposed into a series of approximate motion period units，and R transform characteristics of one approximate motion period unit are extracted.Finally，different types of human action are identified by the method of dynamic time warping.Experimental results show that the proposed method can identify abnormal human behavior effectively.

Key words: activity recognition, approximate motion period analysis, R transform, dynamic time warping

摘要： 从人体行为动作的近似运动周期出发，对人体异常行为识别进行了研究。首先采用混合高斯模型的背景差分法提取出人体运动目标，然后对人体形状的变化进行分析来获取该人体运动的近似周期，将人的行为序列分解为一系列的近似运动周期单元，并提取某一近似运动周期单元的R变换特征，最后通过动态时间规整法来决定不同人体运动的类别归属。实验证明，该算法可以有效地检测人体异常行为。

关键词: 行为识别, 近似运动周期分析, R变换, 动态时间规整