文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Achieving Real-Time Object Detection and Tracking Under Extreme Conditions
侵权投诉

Achieving Real-Time Object Detection and Tracking Under Extreme Conditions

Abstract — In this survey, we present a brief analysis of some of the recent single camera object detection and tracking methods that are developed to function robustly under extreme conditions such as erratic object and camera motion, drastic illuminatio

1

AchievingReal-TimeObjectDetectionandTrackingUnderExtremeConditions

FatihPorikli

Abstract—Inthissurvey,wepresentabriefanalysisofsomeoftherecentsinglecameraobjectdetectionandtrackingmethodsthataredevelopedtofunctionrobustlyunderextremeconditionssuchaserraticobjectandcameramotion,drasticilluminationchanges,andseverenoisecontaminationwhileachievingareal-timeperformance.

I.INTRODUCTION

O

BJECTtrackingisonethemostimportanttasksincomputervision.Invideosurveillance,itisusedtounderstandmovementpatternsofpeopletouncoversuspiciousevents.Itisakeycomponentinreal-timetraf cmanagementtoestimatevehiclemotionstatisticsandcongestionstatus.Advancedvehiclecontrolsystemsdependonthetrackinginformationtokeepthevehicleinlaneandpreventfromcolli-sions.Inmedical eld,trackinghelpsimprovingthequalityoflifeforphysicaltherapypatientsanddisabledpeople.Inretailspaceinstrumentation,itisusedtoadvancearchitecturedesignbyanalyzingtheshoppingbehaviorofcustomers.Inaddition,itprovidescomprehensiblevisualinformationtoattainenvi-ronmentalawarenessinrobotics.Invideosummarization,itisappliedtogenerateobject-basedrepresentationsandautomaticannotationsofthecontent.Trackingisalsoafundamentaltechnologytoextractregionsofinterestandvideoobjectlayersde nedinJPEG-2000andMPEG-4standards.

Asessentialasitisinmanyapplications,robustobjecttrackingunderuncontrolledconditionsstillposesachallenge.Practicalsystemshavetotrackobjectswhenthelightingconditionchangessuddenly,therelativecamera-objectmotionbecomeslarge,thecolorcontrastbecomeslow,andtheimagenoisearisestoanintolerablelevel.Tomakeeverythingworse,thecomputationalcomplexityisrequiredtobekeptataminimumleveltoachievereal-timeperformance.

Intherestofthispaper,webrie ydescribejustafewofthestate-of-artmethodsdesignedtoaddressthesedif culties.

II.OBJECTDETECTION

Backgroundsubtractionisacommonapproachfordiscrim-inatingmovingregionsin xedcamerasetups.Basically,apixel-wisereferencemodelforthestationarypartofthesceneisestimated.Then,theobservedimageiscomparedwiththisreferencetoobtaintheforeground.

Simply,areferenceframecanbecomputedbyaggregatingthepreviousframesinamovingtemporalwindowbyα-blending.Althoughthisapproachhasaminimumcomputa-tionalcost,itinducesghosteffectsandrarelyworksinreal-lifecircumstanceswherethebackgroundisoftennonstationaryduetotheilluminationchanges,shadows,swayingtrees,etc.

Alternatively,predictivetechniquessuchasKalman[1]andWiener[2] ltersareappliedtolearntheunderlyingpixelintensitydistribution.Thesetechniquesstronglydependonthepresetstatetransitionparametersandfailincasethedistributiondoesnot tintoasinglemodelorvariesrandomly.Tohandlesuchmultimodalbackgrounds,mixtureofmodelsthatare exibleenoughtohandlevariationsinlighting,mov-ingsceneclutter,multiplemovingobjectsandotherarbitrarychangestotheobservedscene,areproposed.Inadditiontocolor,therearevariantsthataccommodategradientandoptical owinformation[3].Often,thesemodelsareassignedasGaussianfunctions.Onlineapproximationssuchasexpec-tationmaximization(EM)algorithmareusedtoupdatethemodels[4].However,onlineEMupdatetendstointermingleweakmodesintostrongermodes,thus,distortsthemodelmeansinthelongtermasshowninFig.5.Toadaptthemodelsaccurately,wedevelopedaBayesianupdatemechanism[5]thatcanalsoestimatethenumberofrequiredlayers.Anotherapproachthatmodelsthebackgrounddistributionisthenon-parametrickerneldensityestimation[6].Thismethodkeepsthecolorvaluesofthemultipleframesandestimatesthedensityfunctionusingalltheavailabledatainsteadofonlineapproximation.Proportionaltothenumberofframes,bothmemoryandcomputationalcostsbecomeprohibitivetoadaptthismethodforreal-timeapplications.Althoughthemixtureofmodelsapproachcanconvergetoanyarbitrarydistributionprovidedenoughnumberofcomponents,itscomputationalcomplexityboostsexponentiallyasthenumberofmodelsincreases.

Amajorshortcomingoftheabovemethodsisthattheyallneglectthetemporalcorrelationamongthecolorvalues.Thispreventsthemfromdetectingastructuredorperiodicchange,whichisoftenthecase,sincereal-worldphysicsofteninducesnear-periodicphenomenonintheenvironment:themotionofplantsdrivenbywind,theactionofwavesonabeach,andtheappearanceofrotatingobjects.Todistinguishsuchperiodicmotionfromtheobject’smotion,weproposedafrequencydecompositionbasedrepresentationofthebackground,wave-back[7].Thisalgorithmdetectsnewobjectsbasedsolelyonthedynamicsofthepixelsinasceneratherthantheirap-pearance.Thisisaccomplishedbydirectlyestimatingmodelsofcyclostationaryprocessestoexplaintheobserveddynamicsofthesceneandthencomparingnewobservationsagainstthosemodels.Foragivenframe,wecomputethefrequencycoef cientsandcomparethemtothebackgroundcoef cientstoobtainadistancemapasinFig.4.

Detectionofthetime-varyingphenomenonisalsoattemptedusingcorner-basedbackgroundmodels[8].First,theydetectfeaturepointsusingacornerdetectorandrepresentthem

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email