中文摘要：

通过前景分割和形态学处理得到人体剪影,用链码来表述剪影,求取链码的方向直方图并对其归一化得到一幅剪影图像的稳定特征向量.在识别过程中为了解决行为识别中的时间偏移和速度差异问题,改进了动态时间弯折算法,不使用2个序列的起始端点,而是在模型序列中搜索得到相对于测试序列的局部最优起始端点.最后,对不同行为主体、不同图像尺度的5种人类行为（伸平左臂、挥舞左手、走、坐、下蹲）进行了识别实验,在±30°视角范围内识别率达以95%以上.

英文摘要：

Human silhouette was obtained by foreground segmentation and morphological processing. Then it was descript by a chain code. The direction histogram was calculated from the chain code. Normalizing the chain code direction histogram was the feature vector of a silhouette picture. During the recognition, the dynamic time warping algorithm was improved to deal with the problems of time shifts and activity speed variations. At last, the experiment of five human activities, including raising left arm, waving left arm, sitting, walk, squat was carried on. In the range of ±30°, the recognition rate more than 95 %.