Optical Flow Algorithms

クラス
class	cv::optflow::DualTVL1OpticalFlow
	"Dual TV L1" Optical Flow Algorithm. [詳解]
class	cv::optflow::PCAPrior
	This class can be used for imposing a learned prior on the resulting optical flow. Solution will be regularized according to this prior. You need to generate appropriate prior file with "learn_prior.py" script beforehand. [詳解]
class	cv::optflow::OpticalFlowPCAFlow
	PCAFlow algorithm. [詳解]
class	cv::optflow::RLOFOpticalFlowParameter
	This is used store and set up the parameters of the robust local optical flow (RLOF) algoritm. [詳解]
class	cv::optflow::DenseRLOFOpticalFlow
	Fast dense optical flow computation based on robust local optical flow (RLOF) algorithms and sparse-to-dense interpolation scheme. [詳解]
class	cv::optflow::SparseRLOFOpticalFlow
	Class used for calculation sparse optical flow and feature tracking with robust local optical flow (RLOF) algorithms. [詳解]
struct	cv::optflow::GPCPatchDescriptor
struct	cv::optflow::GPCPatchSample
class	cv::optflow::GPCTrainingSamples
	Class encapsulating training samples. [詳解]
struct	cv::optflow::GPCTrainingParams
	Class encapsulating training parameters. [詳解]
struct	cv::optflow::GPCMatchingParams
	Class encapsulating matching parameters. [詳解]
class	cv::optflow::GPCTree
	Class for individual tree. [詳解]
class	cv::optflow::GPCForest< T >
class	cv::optflow::GPCDetails

型定義
typedef std::vector< GPCPatchSample >	cv::optflow::GPCSamplesVector

列挙型
enum	cv::optflow::SupportRegionType { cv::optflow::SR_FIXED = 0 , cv::optflow::SR_CROSS = 1 }
enum	cv::optflow::SolverType { cv::optflow::ST_STANDART = 0 , cv::optflow::ST_BILINEAR = 1 }
enum	cv::optflow::InterpolationType { cv::optflow::INTERP_GEO = 0 , cv::optflow::INTERP_EPIC = 1 , cv::optflow::INTERP_RIC = 2 }
enum	cv::optflow::GPCDescType { cv::optflow::GPC_DESCRIPTOR_DCT = 0 , cv::optflow::GPC_DESCRIPTOR_WHT }
	Descriptor types for the Global Patch Collider. [詳解]

関数
CV_EXPORTS_W void	cv::optflow::calcOpticalFlowSF (InputArray from, InputArray to, OutputArray flow, int layers, int averaging_block_size, int max_flow)
CV_EXPORTS_W void	cv::optflow::calcOpticalFlowSF (InputArray from, InputArray to, OutputArray flow, int layers, int averaging_block_size, int max_flow, double sigma_dist, double sigma_color, int postprocess_window, double sigma_dist_fix, double sigma_color_fix, double occ_thr, int upscale_averaging_radius, double upscale_sigma_dist, double upscale_sigma_color, double speed_up_thr)
	Calculate an optical flow using "SimpleFlow" algorithm. [詳解]
CV_EXPORTS_W void	cv::optflow::calcOpticalFlowSparseToDense (InputArray from, InputArray to, OutputArray flow, int grid_step=8, int k=128, float sigma=0.05f, bool use_post_proc=true, float fgs_lambda=500.0f, float fgs_sigma=1.5f)
	Fast dense optical flow based on PyrLK sparse matches interpolation. [詳解]
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_DeepFlow ()
	DeepFlow optical flow algorithm implementation. [詳解]
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_SimpleFlow ()
	Additional interface to the SimpleFlow algorithm - calcOpticalFlowSF()
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_Farneback ()
	Additional interface to the Farneback's algorithm - calcOpticalFlowFarneback()
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_SparseToDense ()
	Additional interface to the SparseToDenseFlow algorithm - calcOpticalFlowSparseToDense()
CV_EXPORTS_W Ptr< DualTVL1OpticalFlow >	cv::optflow::createOptFlow_DualTVL1 ()
	Creates instance of cv::DenseOpticalFlow
CV_EXPORTS_W void	cv::motempl::updateMotionHistory (InputArray silhouette, InputOutputArray mhi, double timestamp, double duration)
	Updates the motion history image by a moving silhouette. [詳解]
CV_EXPORTS_W void	cv::motempl::calcMotionGradient (InputArray mhi, OutputArray mask, OutputArray orientation, double delta1, double delta2, int apertureSize=3)
	Calculates a gradient orientation of a motion history image. [詳解]
CV_EXPORTS_W double	cv::motempl::calcGlobalOrientation (InputArray orientation, InputArray mask, InputArray mhi, double timestamp, double duration)
	Calculates a global motion orientation in a selected region. [詳解]
CV_EXPORTS_W void	cv::motempl::segmentMotion (InputArray mhi, OutputArray segmask, CV_OUT std::vector< Rect > &boundingRects, double timestamp, double segThresh)
	Splits a motion history image into a few parts corresponding to separate independent motions (for example, left hand, right hand). [詳解]
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_PCAFlow ()
	Creates an instance of PCAFlow
CV_EXPORTS_W void	cv::optflow::calcOpticalFlowDenseRLOF (InputArray I0, InputArray I1, InputOutputArray flow, Ptr< RLOFOpticalFlowParameter > rlofParam=Ptr< RLOFOpticalFlowParameter >(), float forwardBackwardThreshold=0, Size gridStep=Size(6, 6), InterpolationType interp_type=InterpolationType::INTERP_EPIC, int epicK=128, float epicSigma=0.05f, float epicLambda=100.f, int ricSPSize=15, int ricSLICType=100, bool use_post_proc=true, float fgsLambda=500.0f, float fgsSigma=1.5f, bool use_variational_refinement=false)
	Fast dense optical flow computation based on robust local optical flow (RLOF) algorithms and sparse-to-dense interpolation scheme. [詳解]
CV_EXPORTS_W void	cv::optflow::calcOpticalFlowSparseRLOF (InputArray prevImg, InputArray nextImg, InputArray prevPts, InputOutputArray nextPts, OutputArray status, OutputArray err, Ptr< RLOFOpticalFlowParameter > rlofParam=Ptr< RLOFOpticalFlowParameter >(), float forwardBackwardThreshold=0)
	Calculates fast optical flow for a sparse feature set using the robust local optical flow (RLOF) similar to optflow::calcOpticalFlowPyrLK(). [詳解]
CV_EXPORTS_W Ptr< DenseOpticalFlow >	cv::optflow::createOptFlow_DenseRLOF ()
	Additional interface to the Dense RLOF algorithm - optflow::calcOpticalFlowDenseRLOF()
CV_EXPORTS_W Ptr< SparseOpticalFlow >	cv::optflow::createOptFlow_SparseRLOF ()
	Additional interface to the Sparse RLOF algorithm - optflow::calcOpticalFlowSparseRLOF()
void	cv::optflow::GPCForest< T >::findCorrespondences (InputArray imgFrom, InputArray imgTo, std::vector< std::pair< Point2i, Point2i > > &corr, const GPCMatchingParams params=GPCMatchingParams()) const
	Find correspondences between two images. [詳解]

詳解

Dense optical flow algorithms compute motion for each point:

• cv::optflow::calcOpticalFlowSF
• cv::optflow::createOptFlow_DeepFlow

Motion templates is alternative technique for detecting motion and computing its direction. See samples/motempl.py.

• cv::motempl::updateMotionHistory
• cv::motempl::calcMotionGradient
• cv::motempl::calcGlobalOrientation
• cv::motempl::segmentMotion

Functions reading and writing .flo files in "Middlebury" format, see: http://vision.middlebury.edu/flow/code/flow-code/README.txt

• cv::optflow::readOpticalFlow
• cv::optflow::writeOpticalFlow

列挙型詳解

GPCDescType

enum cv::optflow::GPCDescType

Descriptor types for the Global Patch Collider.

列挙値
GPC_DESCRIPTOR_DCT	Better quality but slow
GPC_DESCRIPTOR_WHT	Worse quality but much faster

InterpolationType

enum cv::optflow::InterpolationType

列挙値
INTERP_GEO	Fast geodesic interpolation, see [Geistert2016]
INTERP_EPIC	Edge-preserving interpolation using ximgproc::EdgeAwareInterpolator, see [Revaud2015],Geistert2016.
INTERP_RIC	SLIC based robust interpolation using ximgproc::RICInterpolator, see [Hu2017].

SolverType

enum cv::optflow::SolverType

列挙値
ST_STANDART	Apply standard iterative refinement
ST_BILINEAR	Apply optimized iterative refinement based bilinear equation solutions as described in [Senst2013]

SupportRegionType

enum cv::optflow::SupportRegionType

列挙値
SR_FIXED	Apply a constant support region
SR_CROSS	Apply a adaptive support region obtained by cross-based segmentation as described in [Senst2014]

関数詳解

calcGlobalOrientation()

CV_EXPORTS_W double cv::motempl::calcGlobalOrientation	(	InputArray	orientation,
		InputArray	mask,
		InputArray	mhi,
		double	timestamp,
		double	duration
	)

Calculates a global motion orientation in a selected region.

引数

orientation	Motion gradient orientation image calculated by the function calcMotionGradient
mask	Mask image. It may be a conjunction of a valid gradient mask, also calculated by calcMotionGradient , and the mask of a region whose direction needs to be calculated.
mhi	Motion history image calculated by updateMotionHistory .
timestamp	Timestamp passed to updateMotionHistory .
duration	Maximum duration of a motion track in milliseconds, passed to updateMotionHistory

The function calculates an average motion direction in the selected region and returns the angle between 0 degrees and 360 degrees. The average direction is computed from the weighted orientation histogram, where a recent motion has a larger weight and the motion occurred in the past has a smaller weight, as recorded in mhi .

calcMotionGradient()

CV_EXPORTS_W void cv::motempl::calcMotionGradient	(	InputArray	mhi,
		OutputArray	mask,
		OutputArray	orientation,
		double	delta1,
		double	delta2,
		int	apertureSize = 3
	)

Calculates a gradient orientation of a motion history image.

引数

mhi	Motion history single-channel floating-point image.
mask	Output mask image that has the type CV_8UC1 and the same size as mhi . Its non-zero elements mark pixels where the motion gradient data is correct.
orientation	Output motion gradient orientation image that has the same type and the same size as mhi . Each pixel of the image is a motion orientation, from 0 to 360 degrees.
delta1	Minimal (or maximal) allowed difference between mhi values within a pixel neighborhood.
delta2	Maximal (or minimal) allowed difference between mhi values within a pixel neighborhood. That is, the function finds the minimum ( ) and maximum ( ) mhi values over neighborhood of each pixel and marks the motion orientation at as valid only if
apertureSize	Aperture size of the Sobel operator.

The function calculates a gradient orientation at each pixel as:

In fact, fastAtan2 and phase are used so that the computed angle is measured in degrees and covers the full range 0..360. Also, the mask is filled to indicate pixels where the computed angle is valid.

覚え書き

• (Python) An example on how to perform a motion template technique can be found at opencv_source_code/samples/python2/motempl.py

calcOpticalFlowDenseRLOF()

CV_EXPORTS_W void cv::optflow::calcOpticalFlowDenseRLOF	(	InputArray	I0,
		InputArray	I1,
		InputOutputArray	flow,
		Ptr< RLOFOpticalFlowParameter >	rlofParam = Ptr< RLOFOpticalFlowParameter >(),
		float	forwardBackwardThreshold = 0,
		Size	gridStep = Size(6, 6),
		InterpolationType	interp_type = InterpolationType::INTERP_EPIC,
		int	epicK = 128,
		float	epicSigma = 0.05f,
		float	epicLambda = 100.f,
		int	ricSPSize = 15,
		int	ricSLICType = 100,
		bool	use_post_proc = true,
		float	fgsLambda = 500.0f,
		float	fgsSigma = 1.5f,
		bool	use_variational_refinement = false
	)

Fast dense optical flow computation based on robust local optical flow (RLOF) algorithms and sparse-to-dense interpolation scheme.

The RLOF is a fast local optical flow approach described in [Senst2012] [Senst2013] [Senst2014] and [Senst2016] similar to the pyramidal iterative Lucas-Kanade method as proposed by [Bouguet00]. More details and experiments can be found in the following thesis [Senst2019]. The implementation is derived from optflow::calcOpticalFlowPyrLK().

The sparse-to-dense interpolation scheme allows for fast computation of dense optical flow using RLOF (see [Geistert2016]). For this scheme the following steps are applied:

1. motion vector seeded at a regular sampled grid are computed. The sparsity of this grid can be configured with setGridStep
2. (optinally) errornous motion vectors are filter based on the forward backward confidence. The threshold can be configured with setForwardBackward. The filter is only applied if the threshold >0 but than the runtime is doubled due to the estimation of the backward flow.
3. Vector field interpolation is applied to the motion vector set to obtain a dense vector field.

引数

I0	first 8-bit input image. If The cross-based RLOF is used (by selecting optflow::RLOFOpticalFlowParameter::supportRegionType = SupportRegionType::SR_CROSS) image has to be a 8-bit 3 channel image.
I1	second 8-bit input image. If The cross-based RLOF is used (by selecting optflow::RLOFOpticalFlowParameter::supportRegionType = SupportRegionType::SR_CROSS) image has to be a 8-bit 3 channel image.
flow	computed flow image that has the same size as I0 and type CV_32FC2.
rlofParam	see optflow::RLOFOpticalFlowParameter
forwardBackwardThreshold	Threshold for the forward backward confidence check. For each grid point a motion vector is computed. If the forward backward error is larger than threshold given by this function then the motion vector will not be used by the following vector field interpolation. denotes the backward flow. Note, the forward backward test will only be applied if the threshold > 0. This may results into a doubled runtime for the motion estimation.
gridStep	Size of the grid to spawn the motion vectors. For each grid point a motion vector is computed. Some motion vectors will be removed due to the forwatd backward threshold (if set >0). The rest will be the base of the vector field interpolation.
interp_type	interpolation method used to compute the dense optical flow. Two interpolation algorithms are supported: INTERP_GEO applies the fast geodesic interpolation, see [Geistert2016]. INTERP_EPIC_RESIDUAL applies the edge-preserving interpolation, see [Revaud2015],Geistert2016.
epicK	see ximgproc::EdgeAwareInterpolator sets the respective parameter.
epicSigma	see ximgproc::EdgeAwareInterpolator sets the respective parameter.
epicLambda	see ximgproc::EdgeAwareInterpolator sets the respective parameter.
ricSPSize	see ximgproc::RICInterpolator sets the respective parameter.
ricSLICType	see ximgproc::RICInterpolator sets the respective parameter.
use_post_proc	enables ximgproc::fastGlobalSmootherFilter() parameter.
fgsLambda	sets the respective ximgproc::fastGlobalSmootherFilter() parameter.
fgsSigma	sets the respective ximgproc::fastGlobalSmootherFilter() parameter.
use_variational_refinement	enables VariationalRefinement

Parameters have been described in [Senst2012], [Senst2013], [Senst2014], [Senst2016]. For the RLOF configuration see optflow::RLOFOpticalFlowParameter for further details.

覚え書き

If the grid size is set to (1,1) and the forward backward threshold <= 0 that the dense optical flow field is purely computed with the RLOF.

SIMD parallelization is only available when compiling with SSE4.1.

Note that in output, if no correspondences are found between I0 and I1, the flow is set to 0.

参照

optflow::DenseRLOFOpticalFlow, optflow::RLOFOpticalFlowParameter

calcOpticalFlowSF() [1/2]

CV_EXPORTS_W void cv::optflow::calcOpticalFlowSF	(	InputArray	from,
		InputArray	to,
		OutputArray	flow,
		int	layers,
		int	averaging_block_size,
		int	max_flow
	)

これはオーバーロードされたメンバ関数です。利便性のために用意されています。元の関数との違いは引き数のみです。

calcOpticalFlowSF() [2/2]

CV_EXPORTS_W void cv::optflow::calcOpticalFlowSF	(	InputArray	from,
		InputArray	to,
		OutputArray	flow,
		int	layers,
		int	averaging_block_size,
		int	max_flow,
		double	sigma_dist,
		double	sigma_color,
		int	postprocess_window,
		double	sigma_dist_fix,
		double	sigma_color_fix,
		double	occ_thr,
		int	upscale_averaging_radius,
		double	upscale_sigma_dist,
		double	upscale_sigma_color,
		double	speed_up_thr
	)

Calculate an optical flow using "SimpleFlow" algorithm.

引数

from	First 8-bit 3-channel image.
to	Second 8-bit 3-channel image of the same size as prev
flow	computed flow image that has the same size as prev and type CV_32FC2
layers	Number of layers
averaging_block_size	Size of block through which we sum up when calculate cost function for pixel
max_flow	maximal flow that we search at each level
sigma_dist	vector smooth spatial sigma parameter
sigma_color	vector smooth color sigma parameter
postprocess_window	window size for postprocess cross bilateral filter
sigma_dist_fix	spatial sigma for postprocess cross bilateralf filter
sigma_color_fix	color sigma for postprocess cross bilateral filter
occ_thr	threshold for detecting occlusions
upscale_averaging_radius	window size for bilateral upscale operation
upscale_sigma_dist	spatial sigma for bilateral upscale operation
upscale_sigma_color	color sigma for bilateral upscale operation
speed_up_thr	threshold to detect point with irregular flow - where flow should be recalculated after upscale

See [Tao2012] . And site of project - http://graphics.berkeley.edu/papers/Tao-SAN-2012-05/.

覚え書き

• An example using the simpleFlow algorithm can be found at samples/simpleflow_demo.cpp

calcOpticalFlowSparseRLOF()

CV_EXPORTS_W void cv::optflow::calcOpticalFlowSparseRLOF	(	InputArray	prevImg,
		InputArray	nextImg,
		InputArray	prevPts,
		InputOutputArray	nextPts,
		OutputArray	status,
		OutputArray	err,
		Ptr< RLOFOpticalFlowParameter >	rlofParam = Ptr< RLOFOpticalFlowParameter >(),
		float	forwardBackwardThreshold = 0
	)

Calculates fast optical flow for a sparse feature set using the robust local optical flow (RLOF) similar to optflow::calcOpticalFlowPyrLK().

The RLOF is a fast local optical flow approach described in [Senst2012] [Senst2013] [Senst2014] and [Senst2016] similar to the pyramidal iterative Lucas-Kanade method as proposed by [Bouguet00]. More details and experiments can be found in the following thesis [Senst2019]. The implementation is derived from optflow::calcOpticalFlowPyrLK().

引数

prevImg	first 8-bit input image. If The cross-based RLOF is used (by selecting optflow::RLOFOpticalFlowParameter::supportRegionType = SupportRegionType::SR_CROSS) image has to be a 8-bit 3 channel image.
nextImg	second 8-bit input image. If The cross-based RLOF is used (by selecting optflow::RLOFOpticalFlowParameter::supportRegionType = SupportRegionType::SR_CROSS) image has to be a 8-bit 3 channel image.
prevPts	vector of 2D points for which the flow needs to be found; point coordinates must be single-precision floating-point numbers.
nextPts	output vector of 2D points (with single-precision floating-point coordinates) containing the calculated new positions of input features in the second image; when optflow::RLOFOpticalFlowParameter::useInitialFlow variable is true the vector must have the same size as in the input and contain the initialization point correspondences.
status	output status vector (of unsigned chars); each element of the vector is set to 1 if the flow for the corresponding features has passed the forward backward check.
err	output vector of errors; each element of the vector is set to the forward backward error for the corresponding feature.
rlofParam	see optflow::RLOFOpticalFlowParameter
forwardBackwardThreshold	Threshold for the forward backward confidence check. If forewardBackwardThreshold <=0 the forward

覚え書き

SIMD parallelization is only available when compiling with SSE4.1.

Parameters have been described in [Senst2012], [Senst2013], [Senst2014] and [Senst2016]. For the RLOF configuration see optflow::RLOFOpticalFlowParameter for further details.

calcOpticalFlowSparseToDense()

CV_EXPORTS_W void cv::optflow::calcOpticalFlowSparseToDense	(	InputArray	from,
		InputArray	to,
		OutputArray	flow,
		int	grid_step = 8,
		int	k = 128,
		float	sigma = 0.05f,
		bool	use_post_proc = true,
		float	fgs_lambda = 500.0f,
		float	fgs_sigma = 1.5f
	)

Fast dense optical flow based on PyrLK sparse matches interpolation.

引数

from	first 8-bit 3-channel or 1-channel image.
to	second 8-bit 3-channel or 1-channel image of the same size as from
flow	computed flow image that has the same size as from and CV_32FC2 type
grid_step	stride used in sparse match computation. Lower values usually result in higher quality but slow down the algorithm.
k	number of nearest-neighbor matches considered, when fitting a locally affine model. Lower values can make the algorithm noticeably faster at the cost of some quality degradation.
sigma	parameter defining how fast the weights decrease in the locally-weighted affine fitting. Higher values can help preserve fine details, lower values can help to get rid of the noise in the output flow.
use_post_proc	defines whether the ximgproc::fastGlobalSmootherFilter() is used for post-processing after interpolation
fgs_lambda	see the respective parameter of the ximgproc::fastGlobalSmootherFilter()
fgs_sigma	see the respective parameter of the ximgproc::fastGlobalSmootherFilter()

createOptFlow_DeepFlow()

CV_EXPORTS_W Ptr< DenseOpticalFlow > cv::optflow::createOptFlow_DeepFlow

(

)

DeepFlow optical flow algorithm implementation.

The class implements the DeepFlow optical flow algorithm described in [Weinzaepfel2013] . See also http://lear.inrialpes.fr/src/deepmatching/ . Parameters - class fields - that may be modified after creating a class instance:

• member float alpha Smoothness assumption weight
• member float delta Color constancy assumption weight
• member float gamma Gradient constancy weight
• member float sigma Gaussian smoothing parameter
• member int minSize Minimal dimension of an image in the pyramid (next, smaller images in the pyramid are generated until one of the dimensions reaches this size)
• member float downscaleFactor Scaling factor in the image pyramid (must be < 1)
• member int fixedPointIterations How many iterations on each level of the pyramid
• member int sorIterations Iterations of Succesive Over-Relaxation (solver)
• member float omega Relaxation factor in SOR

findCorrespondences()

template<int T>

void cv::optflow::GPCForest< T >::findCorrespondences	(	InputArray	imgFrom,
		InputArray	imgTo,
		std::vector< std::pair< Point2i, Point2i > > &	corr,
		const GPCMatchingParams	params = GPCMatchingParams()
	)		const

Find correspondences between two images.

引数

[in]	imgFrom	First image in a sequence.
[in]	imgTo	Second image in a sequence.
[out]	corr	Output vector with pairs of corresponding points.
[in]	params	Additional matching parameters for fine-tuning.

segmentMotion()

CV_EXPORTS_W void cv::motempl::segmentMotion	(	InputArray	mhi,
		OutputArray	segmask,
		CV_OUT std::vector< Rect > &	boundingRects,
		double	timestamp,
		double	segThresh
	)

Splits a motion history image into a few parts corresponding to separate independent motions (for example, left hand, right hand).

引数

mhi	Motion history image.
segmask	Image where the found mask should be stored, single-channel, 32-bit floating-point.
boundingRects	Vector containing ROIs of motion connected components.
timestamp	Current time in milliseconds or other units.
segThresh	Segmentation threshold that is recommended to be equal to the interval between motion history "steps" or greater.

The function finds all of the motion segments and marks them in segmask with individual values (1,2,...). It also computes a vector with ROIs of motion connected components. After that the motion direction for every component can be calculated with calcGlobalOrientation using the extracted mask of the particular component.

updateMotionHistory()

CV_EXPORTS_W void cv::motempl::updateMotionHistory	(	InputArray	silhouette,
		InputOutputArray	mhi,
		double	timestamp,
		double	duration
	)

Updates the motion history image by a moving silhouette.

引数

silhouette	Silhouette mask that has non-zero pixels where the motion occurs.
mhi	Motion history image that is updated by the function (single-channel, 32-bit floating-point).
timestamp	Current time in milliseconds or other units.
duration	Maximal duration of the motion track in the same units as timestamp .

The function updates the motion history image as follows:

That is, MHI pixels where the motion occurs are set to the current timestamp , while the pixels where the motion happened last time a long time ago are cleared.

The function, together with calcMotionGradient and calcGlobalOrientation , implements a motion templates technique described in [Davis97] and [Bradski00] .