摘要
针对现有人体行为识别技术存在的准确率不高和易受环境干扰等缺点,提出一种基于空时特征融合的人体行为识别方法。通过OpenPose提取人体骨骼关节的位置信息用于构造空时融合特征,该特征综合各类行为的空域和时域信息,使得特征表示更具区分度。利用核化主成分分析算法进行特征维度缩减,利用XGBoost算法进行特征分类,获得识别结果。该方法在Multiview Action 3D数据集上进行测试,得到了94.52%的识别率,较现有的其它许多人体行为识别方法表现更好。
Aiming at the problems of low accuracy and anti-disturbance,a body behavior recognition algorithm based on space-time features was proposed.More differentiated space-time features were extracted from skeletal data obtained from OpenPose system,which combined the human postures and temporal order information.The kernel principal component analysis algorithm was used to reduce the dimension of mixed features,and the XGBoost algorithm was employed to classify mixed features to obtain the results.On the MultiView Action 3D data-set,this method scores 94.52%accuracy,which is better than many other existing methods of human behavior recognition.
作者
吕洁
李洪奇
赵艳红
Sikandar Ali
刘艳芳
LYU Jie;LI Hong-qi;ZHAO Yan-hong;Sikandar Ali;LIU Yan-fang(Beijing Key Lab of Petroleum Data Mining,China University of Petroleum(Beijing),Beijing 102249,China;College of Information Science and Engineering,China University of Petroleum(Beijing),Beijing 102249,China)
出处
《计算机工程与设计》
北大核心
2020年第1期246-252,共7页
Computer Engineering and Design
