融合注意力的多残差膨胀三维卷积红外视频行为识别

打开文本图片集

关键词：红外视频；融合注意力；多尺度；行为识别

中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）10-036-3174-08

doi：10.19734/j.issn.1001-3695.2025.01.0030

Infrared video human action recognition based on attention integrated multi-residual inflated 3D ConvNet

Yuan Shuai，Yu Lei†，Wang Jiajie，Xiong Bangshu （KeyLaboratoyfgerocessn&PateRecitioofJgirocencagngongUniesitycgn）

Abstract：Infrared videosarevulnerable tobackground interferenceandlackrichtemporalvariationsduetotheirphysical properties，leading topoorhumanactionrecognition.Toddressthisissue，thispaperproposedamethodcombiningatentionmechanismswithmulti-residualdilated3Donvolutionforinfraredvideoactionrecogition.Fstlyitdesignedtheintegratedulti-imensionalatentio（IMDA）tohandlebackgroundinterferenceandcapturekeyreas inactionframesbyintegratingspatial-temporalandchannelinformation.Secondly，itdesignedthetemporalpyramidsplitatention（TPSA）tocapturedynamicinformationacrossmultiplescalesbysplitingatentioninthetemporal dimension.Basedon TPSA，thispaper furtherdeveloped the temporal pyramidsplitatentionmulti-residual（TSAMR）module toenhancethemodel’sgeneralization inprocessnginfrared videos under diffrent background conditions. Experiments show that this method achieves accuracy of 81. 73% and 90.76% on the IITR-IAR and campus safety action datasets，outperforming existing action recognition methods.

Keywords：infrared video；multi-dimensional attention；multi-scale；action recognition

0 引言

人体行为识别，即依据采集的数据深入分析并准确归类人体的各种动作和姿态，对于众多智能服务而言，是构建其功能的基石。（剩余20032字）