搜索
基于视觉的运动目标检测与跟踪技术研究重庆大学硕士学位论文（学术学位）学生姓名：指导教师：专业：控制科学与工程学科门类：工学重庆大学自动化学院二O一四年四Vision-basedMovingObjectDetection&TrackingTechnologyResearchAThesisSubmittedtoChongqingUniversityInPartialFulfillmentoftheRequirementforMaster’sDegreeofEngineeringBySupervisedbySpecialty:ControlScienceandEngineeringCollegeofAutomationofChongqingUniversity,Chongqing,ChinaApril,2014重庆大学硕士学位论文中文摘要摘要本文主要对基于视觉的运动目标检测与跟踪技术进行初步研究，在研究了单摄像机下的目标跟踪后，分析不足之后，提出采用PTZ摄像机进行改进，同时进一步使用多摄像机协同跟踪的方法扩大跟踪效果。多摄像机协同下的跟踪作为一种优秀的运动目标跟踪算法，长期受到研究学者的关注。复杂的自然环境对背景产生诸多干扰因素，给跟踪造成极大挑战。单个摄像机本身性能的限制，一定程度上影响了跟踪的性能。这些都使得多摄像机跟踪系统的应用与扩展成为必然的事情。本文主要研究了单摄像机和多摄像机下的目标跟踪的研究。对于单摄像机跟踪，提出基于SURF和Kalman滤波的改进算法实现了特定目标的持续稳定跟踪；对于多摄像机协同下的目标跟踪，提出了基于SURF特征模板的交接方法，有效实现了目标交接和目标在的连续跟踪。本文的主要研究工作有：①研究了基于特征点的目标检测技术方法，比较了传统特征点检测算子Harris算子和局部不变性特征检测算子SIFT算子，同时在SIFT家族里面选取了运算速度最快的SURF算子进行比较。通过提取速度和检测精度指标参考下，选择一种算子在两者之间折中的方法为后文跟踪算法的检测方法。这一章主要是为改进算法做前期研究和改进做准备。②传统的SIFT跟踪算法存在的计算量大和跟踪精度偏低的问题，提出了基于区域预测的SURF改进目标跟踪算法。首先用SURF匹配检测出运动目标位置，计算出目标形心，启动Kalman滤波算法进行目标区域预测，在预测区域进行SURF匹配，得到准确的目标位置。为了进一步减少误跟踪率，采用直方图再匹配的方法进行剔错。通过比较与MeanShift算法、模板匹配算法与SIFT跟踪算法的跟踪性能，结果表明，新算法的跟踪性能都明显高于其它几个算法。③本文对多摄像机下的目标协同跟踪的一些关键技术与问题展开了分析研究，着重介绍了基于目标模型的目标交接算法。在本文，采用传递多幅模板图像的方法，这写模板反映目标的多个侧面信息，这样在目标匹配时能够更快定位到目标。在下一摄像机视野内，进行目标匹配，成功率大大提高。实验结果表明，基于SURF特征模型的目标交接算法具有简单实用、准确率高等优点。最终，实现了目标的连续跟踪。在本文的最后，认真总结了本文的重点研究工作，并尝试分析了目标跟踪研究工作的不足之处和可能改进的方向。关键词：目标跟踪；SURF；Kalman滤波；多摄像机；接力跟踪I重庆大学硕士学位论文英文摘要ABSTRACTInthispaper,movingobjectdetectionandtrackingtechnologyhasgotapreliminarystudy.DuetoSomedeficiencyinobjectbasedonsingle-camera,PTZcameraisusedtoimprovethetrackingeffect.Furthermore,multi-cameracoordinationobjecttrackingcanexpandthescopeoftheobjecttracking,whichisadoptedinthispaper.Asanimportantresearchcontentinbothcomputervisionandvideomonitoringfield,multi-cameracoordinationobjecttrackinghasobtainedtheattentionofresearchersforalongtime.Thecomplicatednaturalenvironmentonthebackgroundmakestheobjectfeaturedifficulttoextractandtrack.Also,thelimitedresolutionandthelowscope,alloftheserestricttheobjecttrackingsystemstopromote.Inthestudy,theworkofobjecttrackingbasedonsingle-camera,objecttrackingbasedonPTZcameraandobjecttrackingbasedonmulti-cameracoordinationistobedone.Atthesametime,someimprovedalgorithmsareputforwardtoimprovethetrackingeffect.Inthispaper，themainresearchesareasfollows：①Inthispaper,theobjectdetectionmethodsbasedonfeaturepointshavestudied.Harrisalgorithmisthetraditionalgoodfeaturepointsdetectionalgorithm.Atthesametime,astheemerginglocalinvariantfeatureextractionmethod,SIFTalgorithmiswidelyused.Furthermore,SURFasthespeed-upmethodofSIFTiscomparedwithHarris,SIFT.Aappropriatealgorithmisemployedasthenextimprovedobjecttrackingalgorithmwiththecompromisebetweenextractspeedanddetectionprecision.Thepurposeofthischapteristoserverthenextimprovedalgorithmanddopreliminaryresearch.②HugecomputationandlowtrackingprecisionarethedefectsofthetraditionalSIFT-basedmovingobjecttrackingmethod,whichmaycausetrackingfailure.Tosolvetheseproblems,animprovedmethodbasedonSURFandobjectregionpredictionispresented.SURFisusedtodetecttheobjectpositioninthecurrentframe.Atthen,objectcentriodisobtainedaftercalculatedandthetheobjectregionispredictedbasedontheobjectpositioninformation.Atthesametime,Kalmanfilterisadoptedtopredicttheobjectcentriodinthenextframe.Onlyinthepredictionregion,theobjectisobtainedbySURFtemplatematching.Tofurtherreducetrackingerrorrate,thehistogramre-matchingmethodisused.TheresultsoftheexperimentdemonstratethattheimprovedmethodshowbettertrackingeffectthanMeanShifttracker,templateII重庆大学硕士学位论文英文摘要matchtrackerandSIFTtracker.③Intheresearchofmulti-cameracoordinationobjecttrackingmethod,Somekeytechnologyandproblemshavestudiedinthischapter.Especially,objecthandoveralgorithmbasedonobjecttemplatehasmadethefocus.Inthispaper,theimprovedmethodadoptsthemanytemplateimagetomatchinnextcamera.Moretemplateimage,whichreflectthemultipleobjectprofileinformation,improvethematchingsuccessrate.Theresultsoftheexperimentshowthatthealgorithmhashighaccuracyandsimpleoperation,whichbasedonSURFtemplatemodelforobjecthandover.Keywords：Objecttracking;SURF;Kalmanfilter;Multi-camera;ObjecthandoverIII重庆大学硕士学位论文目录目录摘要········································································································IABSTRACT·································································································II目录······································································································IV第1章绪论··································································································11.1研究背景及意义·······················································································11.2运动目标检测跟踪技术研究现状·································································21.2.1目标特征选取·················································································21.2.2运动目标检测算法···········································································31.2.3运动目标跟踪算法···········································································51.3跟踪技术中的难点及性能要求·········