基于双曲空间的无监督视频异常检测方法

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）07-042-2234-07

doi：10.19734/j. issn.1001-3695.2024.08.0371

Abstract：Inthefieldofvideoanomalydetection，anomalouseventsoftendemonstratetemporalcontinuityandsimlarityExisting unsupervised methods typicallysegment videos into multipleclipsandrandomlyselectsubsets for training，disrupting the continuityofanomalouseventsandcausing thelossofcriticalspatiotemporalinformation.AditionallycurrntEuclidean space-basedmethodsencounterlimitationsinembeddngspacedimensionalitymakingitdificulttoefectivelycapturethelatent geometrichierarchyofvideodata.Toaddress these isues，thispaper introducedanovelunsupervisedvideoanomalydetection methodbasedonhyperbolic space.Itdesignedaspatiotemporalfeatureconstruction（STFC）module toextract temporalcorrelationsand featuresimilaritiesamong videosegments，mbedding themintoLorentzandPoincaréballhyperbolicspaces to learnrichervideorepresentationsthatmoreefectivelydistinguishnormalfromabnormal events.Experimentsshowthatthis method achieves AUC scores of 93.26% and 77.55% on the Shanghai Tech and UCF-Crime datasets，respectively，outperforming existingunsupervised video anomalydetectionmethods.Theseresultsconfirmtheadvantageof hyperbolic spaceincapturingthelatentgeometrichierarchyofvideodataandhighlightitspotential inenhancinganomalydetectioncapabilities.

KeyWords：unsupervised；video anomaly detection；Lorentz hyperbolicspace；Poincaréballhyperbolic space

0 引言

随着国家公共安全意识的不断增强，监控摄像头在街道、十字路口、银行和购物中心等公共场所的使用日益普及，旨在提高整体的公共安全水平。（剩余17085字）