融合时空信息与运动信息的骨架行为识别

打开文本图片集

关键词：骨骼行为识别；PoseConv3D；时空注意力；标签平滑损失函数

中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2025）08-038-2532-06

doi：10.19734/j. issn. 1001-3695.2024. 10.0483

Skeleton-based action recognition through fusion of spatial-temporal and motion information

WeiWei ，Zheng Cheng，Tang Yuan，Li Chen （SchoolofSofware Engineering，Chengdu Universityof Information Technology，Chengdu 61O225，China）

Abstract：Aimingattheproblemof insuffcientutilizationof spatio-temporaldependentfeaturesandmotioninformation inexisting skeletal behaviorrecognition methods，this paper proposedanimproved model thatcombined motion featuredandspatiotemporalatentionbasedonPoseConv3D.Firstly，itusedthelimbheat mapcomposedof limbs，head，andtrunk asinput to activelyenhancethespatialcorrelationofadjacentkeypoints.Secondly，ithancedtheuseofkeyspatio-temporalfeaturesin thebackboneetworkbyitroducingsati-tempral，hanneldmotionexiationmodules.inalyitplacedeoentropylossfunctionwiththelabelsmothinglossfunctiontoimprovethe model’sgeneralizationabilityThemodelachieved recognition accuracies of 94.4% （X-Sub）and 97.5% （X-View）on the NTU RGB+D dataset，and 90.5% （X-Sub）and 91. 4% （X-View）on theNTU RGB +1 D120 dataset. Experiments prove that combining motion features and spatio-temporal attention with skeletal heat maps as input effectively improves the accuracy of behavior recognition.

Key words：skeletonaction recognition；PoseConv3D；spatio-temporal attention；label smoothing loss function

0 引言

人体行为识别是计算机视觉领域的一个重要分支，其在人机交互、视频监控等方面[1～3]发挥着重要作用。（剩余15482字）