基于卷积神经网络的多尺度车辆检测及其域自适应研究

《基于卷积神经网络的多尺度车辆检测及其域自适应研究》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ConvolutionalNeuralNetworkforMulti-scaleVehicleDetectionandDomainAdaptationADissertationSubmittedfortheDegreeofMasterCandidate:XiaoYongjieSupervisor：Prof.XuXuemiaoSouthChinaUniversityofTechnologyGuangzhou,China分类号：TP3学校代号：10561学号：201620131505华南理工大学硕士学位论文基于卷积神

2、经网络的多尺度车辆检测及其域自适应研究作者姓名：肖永杰指导教师姓名、职称：徐雪妙教授申请学位级别：工学硕士学科专业名称：计算机科学与技术研究方向：计算机视觉与图像图形处理论文提交日期：2018年4月20日论文答辩日期：2018年5月31日学位授予单位：华南理工大学学位授予日期：年月日答辩委员会成员：主席：高英委员：李桂清、陈伟能、毕盛、苏锦钿AbstractWiththepopularizationofintelligenttransportationsystem,thedemandofvision-basedv

3、ehicledetectionalgorithmandperformancebecomemoreandmoresevere.Duetothedevelopmentofdeeplearning,convolutional-neural-network(CNN)-basedvehicledetectionapproachesachieveincrediblesuccessinrecentyears.However,existingCNN-basedalgorithmssufferfromtheproblemthatt

4、heconvolutionalfeaturesarescale-sensitiveinobjectdetectiontaskbutitiscommonthattrafficimagesandvideoscontainvehicleswithalargevarianceofscales.Besides,trainingCNNneedsquitealotoftrainingdata.Unfortunately,labelingthesedatasetswillconsumealotofmanpower,resourc

5、esandtime.Astheresultsofthesetwochallenges,itbecomesmoreandmoreimportanttodealwithhowtosolvescale-sensitiveproblemandhowtoconvenientlypreparedatasets.Wedelvedintothechallengesabove.Aimingatthescale-sensitiveproblemofvehicledetection,weproposeascale-insensitiv

6、econvolutionalneuralnetwork(SINet)forfastvehicledetection.Andweproposeadomainadaptationalgorithmfortransferringvehicledetectionmodelindifferentdomains.Aimingattheconvenientlypreparingdatasets,webuildanewhighwaydataset,LSVHdataset.Themaincontentsandcontributio

7、nsofthispaperareasfollows:1.Weproposeascale-insensitiveconvolutionalneuralnetwork(SINet).Afterresearch,werevealtwokeyissues:firstly,existingRoIpoolinglayerintwo-stageobjectdetectionnetworkdestroysthestructureofsmallscaleobjects;Ontheotherhand,thelargeintra-cl

8、assdistanceforalargevarianceofscalesexceedstherepresentationcapabilityofasinglenetwork.Basedonthesetwofindings,wepresentascale-insensitiveconvolutionalneuralnetwork(SINet)forfastdetecting