Motion Computing in Image AnalysisRoadmapImportance of Visual MotionApparent MotionOptic Flow ComputationSlide 6Slide 7Aperture ProblemSlide 9Slide 10Slide 11Optic Flow ConstraintLucas-Kanade Optic FlowSlide 14EdgeLow texture regionHigh texture regionImproving the Lucas-Kanade methodIterative Lucas-Kanade methodSlide 20Slide 21Feature Based MethodSlide 23Estimation CriterionEstimation Criterion(contd.)Slide 26Slide 27Slide 28Slide 29Slide 30Block Matching AlgorithmsBlock Matching Algorithms(contd.)Slide 33Slide 34Slide 35ConclusionMotion Computing in Image Analysis- Mani V ThomasCISC 489/689RoadmapOptic Flow ConstraintOptic Flow ComputationGradient Based ApproachFeature Based ApproachEstimation CriterionBlock Matching algorithmsConclusionSome slides and illustrations are from M. Pollefeys and M. ShahImportance of Visual MotionApparent motion of objects on the image plane is a strong cue to understand structure and 3D motionBiological visual systems infer properties of the 3D world via motionTwo sub-problems of motionProblem of correspondence estimationWhich elements of a frame correspond to which elements of the next frameProblem of reconstructionGiven the correspondence and the camera’s intrinsic parameters can we infer 3D motion and/or structureCourtesy: E. Trucco and A. Verri, “Introductory techniques for 3D Computer Vision”Apparent MotionApparent motion of objects on the image planeCaution required!!Consider a perfectly uniform sphere that is rotating but no change in the light directionOptic flow is zeroPerfectly uniform sphere that is stationary but the light is changingOptic flow existsHope – apparent motion is very close to the actual motionCourtesy: E. Trucco and A. Verri, “Introductory techniques for 3D Computer Vision”Optic Flow ComputationTwo strategies for computing motionDifferential MethodsSpatio temporal derivatives for estimation of flow at every positionMulti-scale analysis required if motion not constrained within a small rangeDense flow measurementsMatching MethodsFeature extraction(Image edges, corners)Feature/Block Matching and error minimizationSparse flow measurementsCourtesy: E. Trucco and A. Verri, “Introductory techniques for 3D Computer Vision”Optic Flow ComputationImage Brightness Constancy assumptionLet E be the image intensity as captured by the cameraUsing Taylor series to expand EApparent brightness of moving objects remains constant ttEyyExxEtyxEttyyxxE ,,,,0dtdEtEdtdyyEdtdxxE tEtyyEtxxELtttyxEttyyxxEL ttt 00,,,,Optic Flow ComputationImage Brightness Constancy assumptionApparent brightness of moving objects remains constantThe are the image gradient while the are the components of the motion fieldCourtesy: E. Trucco and A. Verri, “Introductory techniques for 3D Computer Vision”0tEdtdyyEdtdxxE EyExE , vdtdydtdx , 0tTEE vAperture ProblemWe can measure Terms that can be measuredTerms to be computedNumber of equations - 1The component of the motion field that is orthogonal to the spatial image gradient is not constrained by the image brightness constancy assumptionIntuitivelyThe component of the flow in the gradient direction is determinedThe component of the flow parallel to an edge is unknownCourtesy: E. Trucco and A. Verri, “Introductory techniques for 3D Computer Vision”tEyExE ,,dtdydtdx ,Different physical motion but same measurable motion within a fixed windowRoadmapOptic Flow ConstraintOptic Flow ComputationGradient Based ApproachFeature Based ApproachEstimation CriterionBlock Matching algorithmsConclusionSome slides and illustrations are from M. Pollefeys and M. ShahOptic Flow ConstraintHow to get more equations for a pixel?Basic idea: impose additional constraintsMost common is to assume that the flow field is smooth locallyOne method: pretend the pixel’s neighbors have the same (u,v)If we use a 5x5 window, that gives us 25 equations per pixel! 0. itiEvuE pp 12512225252125252211bdAEEEvuEEEEEEtttyxyxyxpppppppppLucas-Kanade Optic FlowWe now have more equations than unknownsSolve the least squares problemMinimum least squares solution (in d) is given byFirst proposed by Lucas-Kanade in 1981Summation performed over all the pixels in the windowbAdbdA min12512225 tytxyyxyyxxxTTEEEEvuEEEEEEEEbAdAA1252521222Lucas-Kanade Optic FlowLucas-Kanade Optic flowWhen is the Lucas-Kanade equations solvableATA should be invertible ATA should not be too small (effects of noise)Eigenvalues of ATA, 1 and 2 should not be smallATA should be well conditioned1/2 should not be large (1 = larger eigenvalue)tytxyyxyyxxxEEEEvuEEEEEEEEEdgeGradient is large in magnitudeLarge 1 but small 2Low texture regionGradients has small magnitudeSmall 1 and small 2High texture regionGradients are different with large magnitudesLarge 1 and large 2Improving the Lucas-Kanade methodWhen our assumptions are violatedBrightness constancy is not satisfiedThe motion is not smallA point does not move like its neighborsIterative Lucas-Kanade AlgorithmEstimate velocity at each pixel by solving Lucas-Kanade equationsWarp H towards I using the estimated flow fielduse image warping techniquesRepeat until convergenceIterative Lucas-Kanade methodimage Iimage HGaussian pyramid of image H Gaussian pyramid of image Iimage Iimage Hu=10 pixelsu=5 pixelsu=2.5 pixelsu=1.25 pixelsIterative Lucas-Kanade methodimage Iimage JGaussian pyramid of image H Gaussian pyramid of image Iimage Iimage Hrun iterative L-Krun iterative
View Full Document