摘要
Movingobjectextractionandclassificationareimportantproblemsinautomatedvideosurveillancesystems.Abackgroundmodelbasedonregionsegmentationisproposed.AnadaptivesingleGaussianbackgroundmodelisusedinthestableregionwithgradualchanges,andanonparametricmodelisusedinthevariableregionwithjumpingchanges.Ageneralizedagglomerativeschemeisusedtomergethepixelsinthevariableregionandfillinthesmallinterspaces.Atwo-thresholdsequentialalgorithmicschemeisusedtogroupthebackgroundsamplesofthevariableregionintodistinctGaussiandistributionstoacceleratethekerneldensitycomputationspeedofthenonparametricmodel.Inthefeature-basedobjectclassificationphase,thesurveillancesceneisfirstpartitionedaccordingtotheroadboundariesofdifferenttrafficdirectionsandthenre-segmentedaccordingtotheirscenelocalities.Themethodimprovesthediscriminabilityofthefeaturesineachpartition.AdaBoostmethodisappliedtoevaluatetherelativeimportanceofthefeaturesineachpartitionrespectivelyanddistinguishwhetheranobjectisavehicle,asinglehuman,ahumangroup,orabike.Experimentalresultsshowthattheproposedmethodachieveshigherperformanceincomparisonwiththeexistingmethod.
出版日期
2009年04月14日（中国期刊网平台首次上网日期，不代表论文的发表时间）
