资源描述：
《Active Structured Learning for high-speed object detection.pdf》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、ActiveStructuredLearningforHigh-SpeedObjectDetectionChristophH.Lampert,JanPetersMaxPlanckInstituteforBiologicalCyberneticsT•ubingen,Germanyffirstname.lastnameg@tuebingen.mpg.deAbstract.High-speedsmoothandaccuratevisualtrackingofobjectsinarbitrary,unstructuredenvironmentsisessentialforrobotics
2、andhumanmotionanalysis.However,buildingasystemthatcanadapttoarbitraryobjectsandawiderangeoflightingconditionsisachallengingproblem,especiallyifhardreal-timeconstraintsapplylikeinroboticsscenarios.Inthiswork,weintroduceamethodforlearningadiscriminativeobjecttrackingsystembasedontherecentstruct
3、uredregressionframeworkforobjectlocalization.UsingakernelfunctionthatallowsfastevaluationontheGPU,theresultingsystemcanprocessvideostreamsatspeedof100framespersecondormore.Consecutiveframesinhighspeedvideosequencesaretypicallyveryre-dundant,andfortraininganobjectdetectionsystem,itissucientto
4、havetraininglabelsfromonlyasubsetofallimages.Weproposeanactivelearningmethodthatselecttrainingexamplesinadata-drivenway,therebyminimizingtherequirednumberoftraininglabeling.Ex-perimentsonrealisticdatashowthattheactivelearningissuperiortopreviouslyusedmethodsfordatasetsubsamplingforthistask.1I
5、ntroductionSmoothhigh-speedtrackingofarbitraryvisualobjectsisessentialinindustrialautomation,inmanyrobotapplications,e.g.visualservoing,high-speedballgames,andmanipulationofdynamicobjectsincomplexscenarios,aswellasinavarietyofothertopicsrangingfromhumanmotionanalysistoautomaticmicroscopeopera
6、tion.Duetotheimportanceoftheproblem,manydi erentsolutionshavebeenproposedbothinacademicresearchprojectsaswellasforindustrialapplications.Despitegreatprogressincomputervisionresearch,mostsolutionsusedinindustrystillrelyuponcontrolledenvironmentalconditionsthatcanonlybeachievedonfactory
oors.Co
7、mmerciallyavailabletrackingsolutionstypicallyuseactivesolutionssuchaspulsedLED'stargtesorIR-re
ectingmarkers.Ba-sicresearchprojects,ontheotherhand,haveconcentratedeitheroncontrolledsetupswithdarkbackgrounds,oncomplexmarkerpatterns,oronsystems
