为了解决类能量图易受人体运动时间和位置移动等因素影响而难以有效描述动作细节特征的问题, 本文提出了一种基于类能量图金字塔梯度直方图(PHOG)融合特征和多类别Adaboost分类器的人体行为识别方法。该方法首先对经过躯体配准的运动人体目标轮廓图像构造平均运动能量图(AMEI)和增强的运动能量图(EMEI), 分别提取其分层梯度方向直方图(PHOG)特征并进行串联融合, 作为一种多层次的行为特征描述; 然后使用基于查找表的LUT-Real Adaboost算法设计多类别分类器, 实现图像中人体行为动作的识别。实验结果显示其在典型的人体动作数据集DHA上的正确识别率达97.6%, 高于其它采用单一特征描述和SVM等分类器的方法。表明该方法将整体与局部特征相结合, 可以有效描述不同尺度下的动作细节特征, 增强了人体行为特征的描述能力, 提高了识别性能。

In order to solve the problem that energy image species (EIS) are susceptible to human movement time and position shift, i.e., it is difficult to describe the details of human behaviors, in this paper a method of human behavior recognition was present based on pyramid gradient histogram (PHOG) fusion features and a multi-class Adaboost classifier. This method first calculated the average motion energy image (AMEI) and the enhanced motion energy image (EMEI) of an objects silhouette images after human body registration, and then it extracted the PHOG features of AMEI and EMEI and series them together to form a kind of multi-level feature descriptor of human behavior. Finally, a look-up table-based real Adaboost (LUT-Real Adaboost) algorithm was utilized to realize human behavior recognition by designing a multi-class classifier. Experimental results show that the correct recognition rate in typical depth-included human action datasets is 97.6% by using this method, which is higher than that of other classifiers using single feature description and support vector machine. This reveals that, by combining global and local features, the proposed method can effectively describe the detailed active features of human behavior at different scales, enhance the description ability of human behavior characteristics, and improve recognition performance.