关键肢体角度直方图的行为识别  

Action recognition based on the angle histogram of key parts

作　　者：庄伟源 成运[3] 林贤明[1,2] 苏松志[1,2] 曹冬林[1,2] 李绍滋[1,2] 

机构地区：[1]厦门大学信息科学与技术学院,福建厦门361005 [2]福建省仿脑智能系统重点实验室,福建厦门361005 [3]湖南人文科技学院通信与控制工程系,湖北娄底417000

出　　处：《智能系统学报》2015年第1期20-26,共7页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金资助项目(61202143);福建省自然科学基金资助项目(2013J05100;2010J01345;2011J01367);厦门市科技重点项目资助项目(3502Z20123017)

摘　　要：当前的姿态表示的行为识别方法通常对姿态的准确性做了很强的假设,而当姿态分析不精确时,这些现有方法的识别效果不佳。提出了一种低维的、鲁棒的基于关键肢体角度直方图的人体姿态特征描述子,用于将整个动作视频映射成一个特征向量。同时,还在特征向量中引入共生模型,用以表示肢体间的关联性。最后,设计了分层的SVM分类器,第1层主要用于选择高判别力的肢体作为关键肢体,第2层则利用关键肢体的角度直方图并作为特征向量,进行行为识别。实验结果表明,基于关键肢体角度直方图的动作特征具有较好的判别能力,能更好地区分相似动作,并最终取得了更好的识别效果。The current pose-based methods usually make a strong assumption for the accuracy of pose,but when the pose analysis is not precise,these methods cannot achieve satisfying results of recognition. Therefore,this paper proposed a low-dimensional and robust descriptor on the gesture feature of the human body based on the angle histogram of key limbs,which is used to map the entire action video into an feature vector. A co-occurrence model is introduced into the feature vector for expressing the relationship among limbs. Finally,a two-layer support vector machine（ SVM） classifier is designed. The first layer is used to select highly discriminative limbs as key limbs and the second layer takes angle histogram of key limbs as the feature vector for action recognition. Experiment results demonstrated that the action feature based on angle histogram of key limbs has excellent judgment ability,may properly distinguish similar actions and achieve better recognition effect.

关 键 词：角度特征 动作识别 关键肢体 角度直方图 姿态表示 行为分析 动作特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]