基于R-SVM与SVDD的部位外观模型

Part appearance model based on R-SVM and SVDD

为克服现有基于HOG特征的部位外观模型未考虑不同细胞单元的不同作用以及不能准确表征相似度的缺陷,提出了一种基于递归支持向量机(R-SVM)和支持向量数据描述(SVDD)算法的人体部位外观模型。所提外观模型由两个分类器构成,利用R-SVM进行特征选择并建立的分类器用于判断图像某区域是否属于人体部位类,利用SVDD建立的相似度分类器用于计算属于人体部位类的图像区域与外观模型的相似度。将所提部位外观模型用于人体上半身姿态的估计,仿真实验结果显示其比现有部位外观模型的估计准确度更高,表明所提部位外观模型可以更准确地描述真实人体部位。

For overcoming the defect that the existing part appearance models did not consider the different roles of different cells and could not represent the similarity accurately,this paper proposed an appearance model based on the recursive support vector machine（ R-SVM） and support vector data description（ SVDD） algorithm. The proposed appearance model consisted of two classifiers,the classifier built after feature selection by using R-SVM determined whether an image region belonged to the class of human part,the similarity classifier built by using SVDD calculated the similarity of an image region with the proposed appearance model. When used the proposed appearance model to human pose estimation,experiment results show that it can get higher estimation accuracy than the existing part appearance models,that indicate the proposed appearance models can represent real human part more accurately.