资源描述:
《基于时空轨迹行为特征的视频拷贝检测方法》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、xxx计算机研究与发展20xx,xx(x)收稿日期:200x-xx-xx;修回日期:20xx-xx-xx基金资助:国家重点基础研究发展计划项目(973计划,2007CB311100);国家高技术研究发展计划项目(863计划,2007AA01Z416);国家自然科学基金项目(60873165、60802028);北京市科技新星计划项目(2007B071);北京市教育委员会共建项目基于时空轨迹行为特征的视频拷贝检测方法吴潇1,2李锦涛1唐胜1郭俊波11(中国科学院计算技术研究所前瞻研究实验室北京100190)2(中国科学院研究生院北京100049)(wuxiao@ict.ac.cn)Vid
2、eoCopyDetectionBasedonSpatio-TemporalTrajectoryBehaviorFeatureWuXiao1,2,LiJintao1,TangSheng1,andGuoJunbo11(CenterforAdvancedComputingResearch,InstituteofComputingTechnology,ChineseAcademyofSciences,Beijing100190)2(GraduateUniversityoftheChineseAcademyofSciences,Beijing100049)AbstractLargescalevi
3、deocopydetectionistodetectcopiedsegmentsofprovidedvideocontentfromlargevideodatabases.Thisapplicationrequirescompactfeaturewhichisinsensitivetovariousvisualcopychanges.However,traditionalimagefeaturesarepronetospatialchangessuchascolorandtexturetransformations.Thereasonisthecopiedversionshavelar
4、geimagetransformationsincludingglobalqualitydecreaseandlocalvisualdistortions,whichdramaticallychangethedistributionofvisualfeatures.Consequently,previousmethodsbasedonhistogramfeaturesandordinalmeasuresfailincopydetection.Tosolvethisproblem,thispaperproposestheuseofinvariantvisualfeaturesbasedo
5、nkeypointtrajectorybehavior.Insteadofusingthespatialcues,theproposedapproachmodelsthetemporalinformationasrobustfeatures.Temporalcuesarequantifiedbasedonkeypointtrajectorieswhichareinsensitivetostrongchanges.Thenvideosarerepresentedbyspatio-temporalfeatureswhicharemorerobusttocopychanges.BagofTr
6、ajectory(BoT)technicalisadoptedforfastpatternmatchinginlargedatabase.Theexperimentsshowthatspatio-temporaltrajectoryfeaturesarerobusttovariousvisualchanges,includingimageblur,scaleandratiochanges,andminorframeratechange.Comparedwiththestate-of-artschemeusingordinalmeasure,theproposedalgorithmwit
7、hlowercostpresentsbetteraccuracy.Keywordsvideocopydetection;spatio-temporaltrajectory;localfeaturepointdetection;matchingbasedontemporalinfo;Markovstationaryfeature摘要互联网环境中大规模的视频拷贝检测面临拷贝变化多样性问题和数据量大的问题,需要使用鲁棒、精简的视觉特征。提出以视频连续