PUBLIC

LMSOpticFlowC::LMSOpticFlowC(BooleanT) LMSOpticFlowC::SetRegionSize(IntT) LMSOpticFlowC::SetNoiseLevel(RealT) LMSOpticFlowC::SetGradientOrder(IntT) LMSOpticFlowC::SetFilterErode(BooleanT) LMSOpticFlowC::Estimate(const NumImageC &,const NumImageC &) LMSOpticFlowC::Estimate(const PairC> &) LMSOpticFlowC::Motion(void) const LMSOpticFlowC::MotionCov(void) const LMSOpticFlowC::ErrorVar(void) const LMSOpticFlowC::Eigenvalues(void) const LMSOpticFlowC::PrintPS(const FilenameC &,const NumImageC &,RealT,IntT,RealT) const LMSOpticFlowC::PrintPpm(const FilenameC &,const NumImageC &) const

LMSOpticFlowC
Least-mean-squared gradient method for optical flow
include	"amma/Motion/LMSOpticFlow.hh"
User Level:	Default
Library:	LMSMotion
Example:	exLMSOptic.cc
Section:	Image.Motion.Estimation.Optical Flow
In Scope:	std

Comments:
Computes optical flow field from spatial and temporal gradient images using simple least-mean-squared gradient method

The method finds, for each pixel, the least mean squared solution to the motion constraint equation over an image patch (as described here). It also generates a field of covariances of the motion vectors, and a field of variances of the motion constraint equation error. The default parameters are:

 noise (1.0),
 region (9)

RealT noise;

to help bad matrices invert nicely

IntT region;

regression is done over region of this size

BooleanT erode;

if true, fitering operations erode motion field

BooleanT verbose;

IntT grad_order;

order of spatial gradient (1..3)

NumImageC motion;

results of motion estimation

NumImageC cov;

covariance of motion vector estimation

NumImageC var;

variance of motion estimation errors

NumImageC lambda;

eigenvalues of gradient outer product

LMSOpticFlowC(BooleanT Verbose = false)

Constructor

LMSOpticFlowC & SetRegionSize(IntT Size)

Set size of patch over which motion error is minimised

LMSOpticFlowC & SetNoiseLevel(RealT Noise)

Set noise level for matrix inversion

LMSOpticFlowC & SetGradientOrder(IntT Order)

Set order of spatial gradient estimator

LMSOpticFlowC & SetFilterErode(BooleanT Erode)

:Set filter resize to erode (Erode==true) or no erosion (Erode==false) N.B.: even if filters are set for no erosion, there will still be a small amount of erosion of the motion field due to spatial gradient operators

NumImageC<Vector2dC> Estimate(const NumImageC<Vector2dC> & grad,const NumImageC<RealT> & dt)

LMS engine for method; operates on spatial and temporal gradient images

NumImageC<Vector2dC> Estimate(const PairC<NumImageC<RealT>> & im)

Applies method to pair of (filtered) images using first-order differences

NumImageC<Vector2dC> Motion() const

Return field of motion vectors

NumImageC<Matrix2d2C> MotionCov() const

Return field of matrices of motion estimate covariance

NumImageC<RealT> ErrorVar() const

Return field of motion constraint equation error variance

NumImageC<Vector2dC> Eigenvalues() const

Return field of eigenvalues of spatial gradient outer product

void PrintPS(const FilenameC & name = "motion.ps",const NumImageC<RealT> & im = NumImageC<RealT>(),RealT size = 200,IntT subsample = 5,RealT scale = 1) const

Print results in PostScript, using optional image for background

void PrintPpm(const FilenameC & name,const NumImageC<RealT> & im) const

Print results as ppm file: motion vectors are plotted as U/V components of colour image, with 1st original image as Y component