改进后的YOLOv5用于跌倒行为检测张振亚，何明艳，王萍（安徽建筑大学电子与信息工程学院，安徽合肥230022）摘要：随着全球人口老龄化不断加剧，由于跌倒致死的比例也随之增加，及时发现跌倒行为对降低死亡风险至关重要。针对现有跌倒检测算法在实际应用场景中出现漏检、准确率低等问题，本文将改进后的YOLOv5目标检测方法用于跌倒行为检测。具体改进措施：将YOLOv5的边界框损失函数GIoU更换为α-IoU；引入卷积块注意力机制模块（CBAM），使网络可以更专注地学习跌倒特征；在特征融合层引入加权双向特征金字塔网络结构（BiFPN）以充分利用不同尺度的特征，从而提高检测精度。实验结果表明，改进的YOLOv5模型对跌倒行为的检测精度mAP达到了98.8%，比改进前提高了4%，满足对实际应用场景下跌倒检测的要求。关键词：计算机视觉；跌倒检测；YOLOv5；α-IoU；加权双向特征金字塔；卷积块注意力机制中图分类号：TP391.4文献标志码：A文章编号：1007-4260（2023）01-0072-07FallBehaviorDetectionBasedonImprovedYOLOv5ZHANGZhenya,HEMingyan,WANGPing(CollegeofElectronicandInformationEngineering,AnhuiJianzhuUniversity,Hefei230022,China)Abstract:Astheglobalpopulationagingcontinuestointensify,theproportionofhumandeathalsoduetofallsisincreas-ing.Timelydetectionoffallsiscrucialtoreducingtheriskofdeath.Aimingattheproblemsofmissingdetectionandlowaccu-racyofexistingfalldetectionalgorithmsinpracticalapplicationscenarios,thispaperappliestheimprovedYOLOv5targetde-tectionmethodtothefallbehaviordetection.ReplacetheboundingboxlossfunctionGIoUofYOLOv5withα-IoU,thecon-volutionalblockattentionmodule(CBAM)isintroduced,sothatthenetworkcanmorefocusonlearningthefeatureoffalling.Inthefeaturefusionlayer,theweightedbidirectionalfeaturepyramidnetworkstructure(BiFPN)isintroducedtomakefulluseofthefeaturesofdifferentscales,soastoimprovethedetectionaccuracy.Theexperimentalresultsshowthatthedetectionaccuracy(mAP)oftheimprovedYOLOv5modelforfallbehaviorreached98.8%,whichis4%higherthanthatbeforeim-provement,andmeetstherequirementsoffalldetectioninpracticalapplicationscenarios.Keyword:computervision;fallbehaviordetection;Yolov5;α-IoU;BiFPN;CBAM世界老年人口正在迅速增长，我国65岁以...
