opentl::modalities::HistoOrientedGrad Class Reference

Visual modality: Histogram of Oriented Gradients (HOG) This class implements the HOG on a reference pattern, to be matched to the current image It uses the integral HOG in order to speed up computations for multiple evaluations (e.g. particle filters) on a rectangular bounding box at a given pose. More...

Inherits opentl::modalities::Modality.

List of all members.

Public Types

Public Member Functions

virtual HistoOrientedGrad * clone () const

clone the current HistoOrientedGrad class instance

virtual int copyMTrack2MDetect (const TargetPtrVector &targets)

copy corresponding Mtrack features of a state to Mdetect

HistoOrientedGrad (const HistoOrientedGrad &c)

Copy Constructor.

HistoOrientedGrad (opentl::modelprojection::Warp *warp, int camIdx)

Constructor.

virtual void init ()

Needed init() method for parameter handling.

virtual int matchFeatLevel (const TargetPtrVector &targets, T_MEAS_FEATPtrVector &outputMeas, std::size_t partitionIdx)

Matching on feature level

match projected model features with detected image features.

virtual int matchObjLevel (const TargetPtrVector &targets, T_MEAS_OBJPtrVector &outputMeas, std::size_t partitionIdx)

Matching on object level

compute a local, maximum-likelihood estimate of pose (evtl. state), using either pixel- or feature-level likelihood functions.

virtual int preProcess (const opentl::core::cvdata::Image &image)

Compute integral HOG of the current image.

virtual int sampleModelFeatures (const TargetPtrVector &targets)

Sample the HOG from reference image.

virtual int updateModelFeatures (const TargetPtrVector &targets)

Update online reference HOG from the estimated state and the current image.

virtual ~HistoOrientedGrad ()

Destructor.

Protected Member Functions

virtual void computeModelFeatureFromImageGradients (std::vector< double > &MDestination, const IplImage *imgGradx, const IplImage *imgGrady, int offsetx, int offsety, int width, int height)

computes the model features using gradient images for input

virtual void computeModelFeatureFromIntegralHistogram (boost::shared_ptr< opentl::core::State > state, opentl::models::ObjModel &objModel, std::vector< double > &MDestination)

computes the model features using a precomputed integral histogram for input

Detailed Description

Visual modality: Histogram of Oriented Gradients (HOG) This class implements the HOG on a reference pattern, to be matched to the current image It uses the integral HOG in order to speed up computations for multiple evaluations (e.g. particle filters) on a rectangular bounding box at a given pose.

Pixel-level match: (not available) Feature-level match: Bhattacharyya distance between HOGs Object-level match: Brute-force search in discrete state space (N-box)

Author: Giorgio Panin (panin@in.tum.de)

Member Enumeration Documentation

enum opentl::modalities::HistoOrientedGrad::OfflineParameters

Parameters for visual processing in this class.

Enumerator:

paramNcell	Number of histogram cells (default: 9).
OFFLINE_COUNT

Reimplemented from opentl::modalities::Modality.

enum opentl::modalities::HistoOrientedGrad::OnlineParameters

Enumerator:

aperture_size	Sobel filter aperture size (default: 3) Size of the Sobel kernel (must be 1, 3, 5 or 7).
invCovFeat	Inverse variance of residual distance (feature-level).
typeFeatureResidual	Metric of feature-level residual.
matchFeatMdetectWeight	weight for Moffline/Mdetect distance (must be 1-matchFeatMtrackWeight)
matchFeatMtrackWeight	weight for Monline/Mtrack distance (must be 1-matchFeatMdetectWeight)
matchFeatWithMdetect	specify the residual fusion: use Moffline/Mdetect
matchFeatWithMtrack	specify the residual fusion: use Monline/Mtrack
ONLINE_COUNT

Reimplemented from opentl::modalities::Modality.

enum opentl::modalities::HistoOrientedGrad::T_FEATURE_RES

Enumerator:

BHAT
EUCLID

Constructor & Destructor Documentation

opentl::modalities::HistoOrientedGrad::HistoOrientedGrad	(	opentl::modelprojection::Warp *	warp,
		int	camIdx
	)

Constructor.

Parameters:

	warp	pointer to the warp
	camIdx	camera index for this modality

opentl::modalities::HistoOrientedGrad::HistoOrientedGrad ( const HistoOrientedGrad & c )

Copy Constructor.

Parameters:

c

reference HistoOrientedGrad class to copy from

virtual opentl::modalities::HistoOrientedGrad::~HistoOrientedGrad ( ) [virtual]

Destructor.

Member Function Documentation

virtual HistoOrientedGrad* opentl::modalities::HistoOrientedGrad::clone ( ) const [inline, virtual]

clone the current HistoOrientedGrad class instance

Returns:: a pointer to a cloned class instance

Implements opentl::modalities::Modality.

virtual void opentl::modalities::HistoOrientedGrad::computeModelFeatureFromImageGradients	(	std::vector< double > &	MDestination,
		const IplImage *	imgGradx,
		const IplImage *	imgGrady,
		int	offsetx,
		int	offsety,
		int	width,
		int	height
	)			`[protected, virtual]`

computes the model features using gradient images for input

Parameters:

	MDestination	Destination model feature storage container
	imgGradx	Gradient image in x direction
	imgGrady	Gradient image in y direction
	offsetx	position (x) of the lower left corner of the rectangle to sample from
	offsety	position (y) of the lower left corner of the rectangle to sample from
	width	width of the rectangle to sample from
	height	height of the rectangle to sample from

virtual void opentl::modalities::HistoOrientedGrad::computeModelFeatureFromIntegralHistogram	(	boost::shared_ptr< opentl::core::State >	state,
		opentl::models::ObjModel &	objModel,
		std::vector< double > &	MDestination
	)			`[protected, virtual]`

computes the model features using a precomputed integral histogram for input

Parameters:

	state	vector of states to consider for sampling
	MDestination	Destination model feature storage container

virtual int opentl::modalities::HistoOrientedGrad::copyMTrack2MDetect ( const TargetPtrVector & targets ) [virtual]

copy corresponding Mtrack features of a state to Mdetect

Parameters:

states

states vector that should be considered for copying the corresponding features

virtual void opentl::modalities::HistoOrientedGrad::init ( ) [virtual]

Needed init() method for parameter handling.

Reimplemented from opentl::modalities::Modality.

virtual int opentl::modalities::HistoOrientedGrad::matchFeatLevel	(	const TargetPtrVector &	targets,
		T_MEAS_FEATPtrVector &	outputMeas,
		std::size_t	partitionIdx
	)			`[virtual]`

Matching on feature level

match projected model features with detected image features.

Parameters:

	states	Predicted state
	outputMeas	Feature-space measurements and residuals

Reimplemented from opentl::modalities::Modality.

virtual int opentl::modalities::HistoOrientedGrad::matchObjLevel	(	const TargetPtrVector &	targets,
		T_MEAS_OBJPtrVector &	outputMeas,
		std::size_t	partitionIdx
	)			`[inline, virtual]`

Matching on object level

compute a local, maximum-likelihood estimate of pose (evtl. state), using either pixel- or feature-level likelihood functions.

Parameters:

	states	Predicted state
	outputMeas	State-space measurement and residuals (expected state = prediction, observed state = ML estimate)

Reimplemented from opentl::modalities::Modality.

virtual int opentl::modalities::HistoOrientedGrad::preProcess ( const opentl::core::cvdata::Image & image ) [virtual]

Compute integral HOG of the current image.

Parameters:

sensData

input sensor data for preprocessing

virtual int opentl::modalities::HistoOrientedGrad::sampleModelFeatures ( const TargetPtrVector & targets ) [virtual]

Sample the HOG from reference image.

Parameters:

	ref_image	reference image to sample from
	targetId	(optional) specifies the target id inside Mdetect, where the sampled result gets stored Sample the HOG from the shape and appearance model
	states	states vector that should be samples from

Reimplemented from opentl::modalities::Modality.

virtual int opentl::modalities::HistoOrientedGrad::updateModelFeatures ( const TargetPtrVector & targets ) [virtual]

Update online reference HOG from the estimated state and the current image.

Parameters:

states

states vector that should be samples from

Reimplemented from opentl::modalities::Modality.


Public Types
Public Member Functions
virtual HistoOrientedGrad *	clone () const
	clone the current HistoOrientedGrad class instance
virtual int	copyMTrack2MDetect (const TargetPtrVector &targets)
	copy corresponding Mtrack features of a state to Mdetect
	HistoOrientedGrad (const HistoOrientedGrad &c)
	Copy Constructor.
	HistoOrientedGrad (opentl::modelprojection::Warp *warp, int camIdx)
	Constructor.
virtual void	init ()
	Needed init() method for parameter handling.
virtual int	matchFeatLevel (const TargetPtrVector &targets, T_MEAS_FEATPtrVector &outputMeas, std::size_t partitionIdx)
	Matching on feature level match projected model features with detected image features.
virtual int	matchObjLevel (const TargetPtrVector &targets, T_MEAS_OBJPtrVector &outputMeas, std::size_t partitionIdx)
	Matching on object level compute a local, maximum-likelihood estimate of pose (evtl. state), using either pixel- or feature-level likelihood functions.
virtual int	preProcess (const opentl::core::cvdata::Image &image)
	Compute integral HOG of the current image.
virtual int	sampleModelFeatures (const TargetPtrVector &targets)
	Sample the HOG from reference image.
virtual int	updateModelFeatures (const TargetPtrVector &targets)
	Update online reference HOG from the estimated state and the current image.
virtual	~HistoOrientedGrad ()
	Destructor.
Protected Member Functions
virtual void	computeModelFeatureFromImageGradients (std::vector< double > &MDestination, const IplImage imgGradx, const IplImage imgGrady, int offsetx, int offsety, int width, int height)
	computes the model features using gradient images for input
virtual void	computeModelFeatureFromIntegralHistogram (boost::shared_ptr< opentl::core::State > state, opentl::models::ObjModel &objModel, std::vector< double > &MDestination)
	computes the model features using a precomputed integral histogram for input