摘要
由于现有的动作识别方法不能直接用于人体微反应动作识别,本文基于人体下肢微反应动作特点,构建了一种时空金字塔韦伯局部描述子并设计了基于字典学习的人体下肢微反应自动识别算法。该方法利用时空金字塔韦伯局部描述子提取每一类人体下肢微反应动作特征,使用主成分分析法对特征降维;然后,建立每一类动作子字典并将子字典串联形成总的动作字典;最后,通过实验分析了金字塔级数L,降维后每类动作特征维数d_(PCA),每类动作子字典原子个数n_(Atom),以及稀疏阈值C等参数对识别结果的影响,并确定最优参数值L=3,d_(PCA)=30,n_(Atom)=40,C=10。实验结果表明,提出的算法对10种人体下肢微反应动作的识别率均在0.83~0.91之间,平均识别率达到0.86,高于其他动作识别算法。设计的算法更适用于人体下肢微反应动作分类,并可有效提高分类识别率。
Because the existing motion recognition method couldnot be directly used in human micro expression action recognition.A spatio-temporal pyramid Weber Local Descriptor(STPWLD)was constructed and an automatic recognition algorithm of human lower limb micro-expression action based on dictionary learning according to characterize human lower limb micro expression action was designed.With the method,the features of human lower limb micro-expression action was extracted by the STPWLD.Then,the dimensions of STPWLD feature were reduced by the principal component analysis.Furthermore,the sub-dictionaries of human lower limb micro-expressions action was established and these sub-dictionaries were connected in series to construct a general action dictionary.Finally,the influence of the parameters of the algorithm on the recognition results was analyzed,and the optimal value of these parameters was determined.It shows that the optimal value of pyramid scales is3,the optimal feature dimension of each action after dimension reduction is30,the optimal number of atoms in each action dictionary is40and the optimal value of sparse threshold is10.The experimental results indicate that the recognition rates of the proposed algorithm for10kinds of human lower limb micro expression actions are all between0.83~0.91,and the average recognition rate is0.86,higher than that of other algorithms.The algorithm is suitable for the classification of human lower limb micro expression actions and improves the classification recognition rate effectively.
作者
王昊鹏
冯显英
张明亮
WANG Hao-peng;FENG Xian-ying;ZHANG Ming-liang(Shandong Police College, Jinan 250014, China;Key Laboratory of High Efficiency and Clean Mechanical Manufacture, Ministry of Education, Shandong University, Jinan 250061, China;Shandong Academy of Governance, Jinan 250014, China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2017年第11期2947-2957,共11页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.51405194)
山东省自然科学基金资助项目(No.ZR2016EEP13
No.ZR2013CQ025)
山东省高等学校科技计划资助项目(No.J16LN16)
关键词
人体下肢
应激微反应
动作识别
自动识别
时空金字塔韦伯局部描述子
human lower limb
micro expression
action recognition
spatio temporal pyramid weber local descriptor
dictionary learning
