摘要: |
针对长时间动作识别难以充分利用时空域信息的问题,提出了基于相对骨骼点特征和时 序自适应感受野的动作识别方法。首先,该方法在特征获取部分增加了相对骨骼点特征,以满足 节点多样性和互补性要求,将其分别输入到空域图卷积网络,获得空间中相邻关节聚合的局部特 征。然后,设计了一个时序自适应感受野网络,以获取在时域中关节变化的局部特征,并且增加了 网络对不同持续时长动作的适应性。最后,经过决策级融合模块,计算类别概率,得到分类结果。 仿真结果表明,基于 NTU RGB+D 和 Kinetics-skeleton两大基准数据集,对比多种主流方法,均 取得了更高的识别准确率,分别为96.2%与60.1%。该方法可以较好地提取不同动作的区别性时 间特征,提高了动作时空特征的判别能力。 |
关键词: 动作识别 时序特征提取 图卷积网络 相对骨骼点特征 时序自适应 |
DOI: |
|
基金项目:中国国家铁路集团有限公司系统性重大课题(P2020T002) |
|
Action Recognition Based on Relative Skeleton Point Features and Temporal Adaptive Receptive Field |
HU Hao,SHITian-yun,SONGYong-hong,YU Huai |
(PostgraduateDepartment,ChinaAcademyofRailwaySciences,Beijing100081,China;;ChinaAcademyofRailwaySciencesCorporationLimited,Beijing100081,China;;SchoolofSoftware,Xi’anJiaotongUniversity,Xi’an710049,China;;Signal& Communication ResearchInstitute,ChinaAcademyofRailwaySciencesCorporationLimited,Beijing100081,Chin) |
Abstract: |
Aimingattheproblemthatitisdifficulttomakefulluseoftemporalandspatialdomain informationforlong-termactionrecognition,anactionrecognitionmethodbasedonrelativeskele- tonpointfeaturesandtemporaladaptivereceptivefieldisproposed.Firstly,thismethodaddsrel- ativeskeletonpointfeaturesinthefeatureacquisitionparttomeetthenodediversityandcomple- mentarityrequirements,andrespectivelyinputsthemintothespatialgraphconvolutionnetwork toobtainthelocalfeaturesoftheadjacentjointaggregationinspace.Then,atime-seriesadaptive receptivefieldnetworkisdesignedtoobtainthelocalfeaturesofjointchangesinthetimedomain,andtoincreasetheadaptabilityofthenetworktoactionsofdifferentdurations.Finally,through thedecision-levelfusionmodule,thecategoryprobabilityiscalculatedtoobtaintheclassification result. ThesimulationresultsshowthatbasedonthetwobenchmarkdatasetsofNTU RGB+D andKinetics-skeleton,comparedwithothermainstream methods,thismethodhasachievedhigher recognitionaccuracies,whichare96.2%and60.1%,respectively.Themethodcanbetterextract thediscriminativetemporalfeaturesofdifferentactions,andimprovethediscriminationabilityof spatiotemporalactionfeatures. |
Key words: Actionrecognition Temporalfeatureextraction Graphconvolutionalnetwork Rela- tiveskeletonpointfeatures Temporaladaptation |