基于多描述子特征编码人体行为识别.PDFVIP

第28卷 第8期 计 算机 技 术 与发 展 Vol.28 No.8 2018年8月 COMPUTERTECHNOLOGY AND DEVELOPMENT Aug. 2018 基于多描述子特征编码的人体行为识别 宋相法,姚 旭 (河南大学计算机与信息工程学院,河南开封 475004) 摘 要:针对采用单一描述子和单一特征编码方法导致三维人体骨架序列的行为识别率较低的问题,提出一种基于多描 述子特征编码的方法。 首先,从三维人体骨架序列中分别提取运动姿态描述子和角度描述子。 然后,对每种描述子分别 进行向量量化编码、稀疏编码和局部约束线性编码,从而获得六种特征。 最后,根据这六种特征分别构造线性分类器,通 过投票机制得到最终的识别结果。 为了验证所提方法的有效性,在三维人体骨架序列行为数据集MSR Action3D上进行 了实验,实验结果表明该方法的识别率为94.9%,并且高于其他方法的识别率。 关键词:人体行为识别;特征编码;运动姿态描述子;角度描述子;三维人体骨架序列 中图分类号:TP391 文献标识码:A 文章编号:1673-629X(2018)08-0017-05 doi:10.3969/j.issn.1673-629X.2018.08.004 Human Activity Recognition Based on Multi-descriptor Feature Coding SONG Xiang-fa,YAO Xu (School of Computer and Information Engineering,Henan University,Kaifeng475004,China) Abstract:Aiming at theproblemoflowactivityrecognitionratefrom3Dhumanskeletonsequencebasedonasingledescriptorandasin- gle featurecodingmethod,weproposeamethodbased on multi-descriptor featurecoding. Firstly,movingpose descriptor andangle de- scriptor are extracted respectively from3D human skeleton sequence. Then,vector quantization coding,sparse coding and locality con- strained linear coding are employedrespectivelytoget sixkindsoffeaturebasedontwokindsofdescriptor. Finally,linear classifiersare respectively constructed based on these six kindsof feature,and the recognition result isdecided by voting strategy. In order to validate the effectsof the proposed method,the experiment on MSR Action3D,a public 3D human skeleton sequence activity database,demon- stratesthat the proposed method achieves94.9% of recognition accuracy,which is superior to the state-of-art of methods. Key words:human activity recognition;feature coding