基于轻量化神经网络的社交距离检测①王林,张江涛(西安理工大学自动化与信息学院,西安710048)通信作者:张江涛,E-mail:1307510426@qq.com摘要:保持安全社交距离是有效防止病毒传播的重要手段之一,不仅可以减少感染者数量和医疗负担,同时也极大降低死亡率.在YOLOv4框架基础上使用轻量化网络E-GhostNet代替原网络中的CSPDarknet-53,E-GhostNet网络在输入数据和原始Ghost模块生成的输出特征之间建立关系,使网络能够捕获上下文特征.然后,在E-GhostNet中引入坐标注意力机制(CA)增强模型对有效特征的关注.另外,使用SIoU损失函数更换CIoU损失获得更快的收敛速度和优化效果.最后,结合DeepSORT多目标跟踪算法来检测和标记行人,并使用仿射变换(IPM)判定行人间距离的违规行为.实验结果显示,该网络检测速度为40FPS,精度值达到85.71%,相比原始GhostNet算法提升2.57%,达到实时行人距离检测的效果.关键词:YOLOv4;DeepSORT;社交距离;E-GhostNet;轻量化网络;目标检测引用格式:王林,张江涛.基于轻量化神经网络的社交距离检测.计算机系统应用,2023,32(2):128–138.http://www.c-s-a.org.cn/1003-3254/8942.htmlSocialDistanceDetectionBasedonLightweightNeuralNetworkWANGLin,ZHANGJiang-Tao(SchoolofAutomationandInformation,Xi’anUniversityofTechnology,Xi’an710048,China)Abstract:Maintainingasafesocialdistanceisoneoftheimportantmeanstoeffectivelypreventthespreadofthevirus.Moreover,itcannotonlyreducethenumberofinfectedpeopleandeasethemedicalburdenbutalsogreatlylowerthemortalityrate.Onthebasisoftheyouonlylookonceversion4(YOLOv4)framework,thelightweightnetworkE-GhostNetisusedtoreplacetheCSPDarknet-53intheoriginalnetwork.TheE-GhostNetnetworkestablishesarelationshipbetweentheinputdataandtheoutputfeaturesgeneratedbytheoriginalGhostmodule,therebyenablingthenetworktocapturecontextualfeatures.Then,thecoordinateattention(CA)mechanismisintroducedtoE-GhostNettoenhancethemodel’sattentiononeffectivefeatures.Inaddition,thecompleteintersectionoverunion(CIoU)lossfunctionisreplacedbythesoftintersectionoverunion(SIoU)lossfunctiontoobtainafasterconvergencespeedandoptimizationeffect.Finally,theDeepSORTmulti-targettrackingalgorithmisutilizedtodetectandlabelpedestrians,andaffinetransformation(IPM...
