雙流特征增強與融合的弱監(jiān)督時序動作定位
打開文本圖片集
關(guān)鍵詞：弱監(jiān)督；時序動作定位；空洞卷積；雙流融合
中圖分類號：TP391.4 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）07-039-2213-07
doi：10.19734/j. issn.1001-3695.2024.09.0373
Abstract：Weakly supervised temporalaction localizationaims to clasifyand locateaction instances inuntrimmed videos usingonlyvideo-levellabels.Existing models typicallusepre-trainedfeature extractors toextractsegment-levelRGBandopticalflowfeaturesfromvideos，butthepre-extractedsegment-levelvideofeaturesonlycovershorttimespansanddootonsider thecomplementarityandcorrelationbetweeRGBandopticalfo，hichaffctstheacuracyoflocalization.Totisnd，tis paper proposedaweakly-supervised temporalaction localization model withdual-streamfeatureenancementand fusion.Firstly，itexpandedthereceptivefieldthrough amulti-scaledensedilatedconvolution，alowing the modeltocover multipletime spansand capture the temporal dependenciesbetween video segments，resulting inenhancedRGBandopticalflow features. Then，itutilizedaconvolutionalnetworktoadaptivelyextractkeyfeaturesfromtheenhancedRGBandopticalflowfeaturesfor fusion，achievingcomplementarycorelationetweenRGBandopticalflowfeatures，furtherenrichingthevideofeaturerepre sentationand improving theaccuracyof themodel'slocalizationperformance.The modelachieves detectionaccuraciesof （20 73.9% and 43.5% on the THUMOS14 and ActivityNet1.3 datasets respectively，outperforming the existing state-of-the-art models，which proves the effectiveness of the proposed model.
Key Words：weakly supervision；temporal action localization；dilated convolution；dual-stream fusion
0 引言
時序動作定位作為視頻理解領(lǐng)域中的關(guān)鍵任務(wù)，在諸如視頻監(jiān)控、異常檢測和視頻檢索等實際應(yīng)用中扮演著重要角色。（剩余19775字）