通过研究,建立了一个基于主元分析的识别人体行为的系统.其方法是通过在H、S、I颜色空间对皮肤颜色建立高斯模型,结合运动限制和区域连续性,系统地分割并跟踪人脸和双手.然后,在PCA框架下,表示脸和手的运动参数曲线,并和范例进行匹配,这种通过对行为在时空域变化的建模方法,能在行为主体和成象条件有变化的情况下识别行为.以太极拳式为例,来验证方法和系统的效果,实验结果证明了此方法误识率低,有一定的鲁棒性,可应用于建立基于运动语义识别的视频检索、高效视频编码、自动教练等场合.

This paper presents an approach to recognize human activity based on principal component analysis (PCA). By a Gaussian model of skin color in HSI color space, our system tracks human face and hands incorporating with the constraints of motion and region continuity. A constant acceleration motion estimation and Schwarz representation based shape matching are applied to create correspondence between frames. Then motion parameters of faces and hands are represented and matched with the parameter curves of exemplars in PCA framework. Through modeling the spatio-temporal variants of each type of activity, recognition can be achieved although subject and imaging condition are different from those of exemplars. Examples of Taiji postures recognition are studied and discussed to illustrate our method. The experiment shows that this activity recognition approach is of low confusion rate and robust in some degree. We believe this approach can be applied to develop an activity interpretation system.Applications fields of this work include indexing video based on motion semantic description, assistant exercise training, video surveillance efficient video coding, etc.