基于深度运动图的人体行为识别  

Human action recognition based on depth motion maps

作　　者：史东承[1] 李延林[1] 

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130012

出　　处：《长春工业大学学报》2017年第3期276-281,共6页Journal of Changchun University of Technology

基　　金：吉林省教育厅"十三五"规划项目(吉教科合字[2016]第349号)

摘　　要：将人体行为深度映射图(depth map)连续投影到3个互相垂直的笛卡尔平面,然后对投影做绝对差分,累积各自投影面的差分图像,得到完整的人体行为三维信息——深度运动图(Depth Motion Maps,DMMs)。利用MSRAction dataset和3DAction Pairs dataset进行训练以获取人体行为字典。在识别未知动作时,利用Tikhonov矩阵计算得出权重系数向量。最后,利用L2范式正则化协同表示对待识别动作进行分类。通过上述两个数据库的验证,分别达到了95.3%和83.8%的平均识别率,已经达到对DMMs的较高识别率。Human action depth maps are projected continuously to three perpendicular Descartes plane.The projections are absolute differentiated and cumulated to obtain complete 3D information of human action which is called the Depth Motion Maps（DMMs）.With MSRAction dataset and 3DAction Pairs software,DMMs is trained to get the human action dictionary.When the unidentified human action is input,weight coefficients are calculated by using the Tikhonov matrix,and then L2-regularized collaborative representation classifier is used to classify the actions.Two data-set experiments indicate that the average recognition rates is 95.3% and 83.8% respectively.

关 键 词：人体行为 识别 深度运动图 L2范式 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]