基于时空域Adaboost算法的人体动作识别研究  

Human action recognition based on time-space domain Adaboost algorithm

作　　者：郭丹丹[1] 朱希安[1] 

机构地区：[1]北京信息科技大学信息与通信工程学院,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2015年第2期50-54,共5页Journal of Beijing Information Science and Technology University

基　　金：国家科技重大专项(2011ZX05039-004-02);北京信息科技大学研究生科技创新项目(5121424102)

摘　　要：基于Kinect深度传感器提取人体骨骼信息,提出了一种有效的动作识别方法。首先利用Kinect传感器获取人体骨骼中20个骨骼点位置信息和旋转角度信息,用四元数的形式来表示,提取一个时间段的每一个骨骼点的四元数参数,作为80个时空特征;用支持向量机(Support Vector Machine,SVM)对每一个时空特征进行训练,得到80个弱分类器;最后利用Adaboost算法进行训练,得到相应的动作特征集和强分类器,并对强分类器进行融合。实验结果证明,该方法对不同体型的人的动作都具有很高的识别准确率和鲁棒性。An effective human action recognition method based on the human skeletal information extracted by Kinect depth sensor is proposed in this paper. First, 20 of three-dimensional position information of human bones and the rotation information of skeletal point in the form of quaternion are acquired by Kinect depth sensor. Then each of parameters of quaternion is extracted as 80 temporal- spatial characteristics. SVM algorithm are used to train the weak classifier. Finally, a corresponding set of operation characteristics and the strong classifiers are obtained by training samples based on Adaboost algorithm, and the strong classifiers are fused. The experimental results show that the method has a high identification accuracy and robustness for body movements of people with different body features.

关 键 词：Kinect深度传感器 ADABOOST算法 四元数 时空特征 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]