基于位置敏感的TimeSformer的群养猪攻击行为识别

打开文本图片集

中图分类号：S126；TP391.4 文献标识码：A 文章编号：2095-5553（2025）12-0121-07

Abstract：The aggressivebehavior pigs is closelyrelated totheir healthstatus， timelyrecognition pigaggression is crucial formodernbreeding industries.Presently，studiesonaggressionrecognition forpigsare mainlydivided nto2 categories group-levelpairwise-level.However，thispaperattempts turtherrefineaggressionrecognitionforpigs froman individual-levelperspectivethus proposes analgorithm therecognitionaggresivebehaviors in ndividual pigs basedonlocation sensitive TimeSformer.The video8pigs in the pen was recorded for 3 days then 600 aggresive video episodes each lasting1s were labelled.1144 aggressve subvideos3328non-aggressivesub-videos for each pig were segmented from these video episodes.To balance the number positive negative samples，the aggressivesub-videos were exped to3432 byusing horizontal vertical mirroring，resulting in atotal6760 sub-videos as the dataset.This dataset was divided into training，validation， test sets at a ratio . Since aggresive behavior pigs isacontinuous spatial-temporal motion，TimeSformer network wasused to extract the spatial-temporalfeaturespigs.Then，the TransformerEncodermoduleFFN（feed forwardnet）network wereused toreconstructthefeatures.According tothecharacteristicaggesivepigs withcontinuous interactivepositionalchanges， locationsensitivity wasintegrated into TimeSformernetwork tooptimize discrimination features.Finaly，the reconstructed features werepassed throughafullyconnected layertorecognizeaggresivebehavior pigs.Therecognition accuracy this method was 96.4% . The result indicates that this method can be used to efectively recognize aggressive behavior individual pigs.

Keywords：group-housedpigs；aggresive behaviorrecognition；deep learning；livestock poultrybreeding; location sensitivity

0 引言

随着现代猪场养殖规模的扩大化和养殖环境的复杂化，猪的攻击行为监控已备受关注。（剩余11447字）