Real-time image sequence segmentation using curve evolution

RealTimeImageSequenceSegmentationUsingCurveEvolutionJ.Zhang andW.LiuDepartment of Electrical Engineering and ComputerScienceUniversityof Wisconsin-MilwaukeeMilwaukee,WI53201junzhang@uwm.edu,wliu@uwm.eduAbstractIn this pap er, we describ e a novel approach to image sequence segmentation and its real-timeimplementation.This approach uses the 3D structure tensor to pro duce a more robust framedi erencesignalandusescurveevolutiontoextractwhole(moving)ob jects.Ouralgorithmis implemented on a standard PC running the Windows op erating system with video capturefromaUSBcamerathatisstandardWindowsvideocapturedevice.UsingthestandardvideoI/Ofunctionalities,oursegmentationsoftwareishighlyp ortableandeasytomaintainandupgrade.InitscurrentimplementationonaPentium400,thesystemcanp erform segmentation at 5 frames/sec with a frame resolution of 160120.I.Intro ductionSeparatingmovingob jectsfromthebackgroundinavideoclipisknownasimagese-quencesegmentationproblem.Inrecentyears,ithasattractedconsiderableinterestduetoits applications in a wide range ofareas.These include ob ject-basedvideo compression (e.g.,MPEG4), video surveillance and monitoring, automatic target detection and tracking, and theanalysis of medical and other scienti c image sequences.Most previous work in image sequence segmentation can b e classi ed into three approaches:1)intra-frame segmentation,2) motion eld clustering/segmentation, and 3) frame di erenc-ing.In the rstapproach(e.g.,see[1 ]),eacframein animagesequenceis segmentedinde-p endently (i.e., intra-frame) into regions of homogeneous intensity or texture, using traditionalimage segmentationtechniques.Then regions in consecutive frames are matched and tracked.This would work well if intra-frame segmentation pro duced a small numb er of regions closelyrelatedtoreal-worldob jects.However,sinceimagesegmentationusingintensityortextureremainsadicultproblem,thisapproachoftenencountersdiculties,asmanyimageseg-mentation techniques pro duce over-segmentation.In the second approach, a dense motion eld is used for segmentation.For example, pixelswithsimilarmotionvectorsaregroup edintoregions(e.g.,see[2]-[3 ]).Alternatively,smallThis work is supp orted in part by ONR grant 00PR01565-00.