说明：  针对行为识别中行为者朝向变化带来的问题,提出了一种基于人体行为3D模型的2D行为识别算法.在学习行为 分类器时,以3D占据网格表示行为样本,提取人体3D关节点作为描述行为的特征,为每一类行为训练一个基于范例的隐马 尔可夫模型(Exemplar-based hidden Markov model,EHMM),同时从3D行为样本中选取若干帧作为3D关键姿势集,这个 集合是连接2D观测样本和人体3D关节点特征的桥梁.在识别2D行为时,2D观测样本序列可以由一个或多个非标定的摄 像机采集.首先在3D关键姿势集中为每一帧2D观测样本寻找与之最匹配的3D关键姿势帧,之后由行为分类器对2D观测 样本序列对应的3D关键姿势序列进行识别.该算法在训练行为分类器时要进行行为者的3D重构和人体3D关节点的提取,
(Identification of actors for the behavior problems caused by changes in direction, a 3D model of human behavior based on 2D behavior recognition algorithms. In learning behavior Classifier when the 3D mesh that acts of take samples of 3D joint points extracted human behavior characteristics as described for each type of sample-based behavioral training a hidden Markov Seoul Markov model (Exemplar-based hidden Markov model, EHMM), while selected samples from the 3D behavior of a number of key frames as a 3D position set, the Observed sample collection is connected 2D and 3D joint points of human feature of the bridge. 2D behavior in the identification, 2D sample sequence can be observed by one or more non-calibrated camera Camera acquisition. First of all key positions in the 3D 2D focus for each frame with which the observed sample to find the best match of the 3D key frame position, followed by the classifier behavior observed on 2D Sample sequence corresponding to the 3D se)