#include <MeanShiftTracker.h>

Public Types
enum	kernelType { epanechnikov, gauss }
	An enumeration for the different kernel types. More...

Public Member Functions
	MeanShiftTracker (kernelType type, int bandwidth, float stdDev=1)
	Constructor with only the most needed parameters. More...

void	setKernel (kernelType type, int bandwidth, float stdDev=1)
	sets new kernel type (with params) More...

int	getKernel () const
	returns current kernel type More...

const core::Img32f &	getKernelImage () const
	returns current kernel image More...

int	getBandwidth () const
	Returns current kernel bandwidth. More...

const utils::Point32f	step (const core::Img32f &weigthImage, const utils::Point32f &initialPoint, int maxCycles=-1, float convergenceCriterion=1.0, bool *converged=0)
	This function returns a new center after the MeanShift algorithm is applied. More...

Static Public Member Functions
static core::Img32f	generateEpanechnikov (int bandwidth)
	Generates an Epanechnikov Kernel. More...

static core::Img32f	generateGauss (int bandwidth, float stdDev)
	Generates a Gauss Kernel. More...

Private Member Functions
utils::Point32f	applyMeanShiftStep (const core::Img32f &image, const utils::Point32f &pos)
	Applies a single step of the mean shift algorithm. More...

Private Attributes
int	m_bandwidth
	The bandwidth of the kernel; equals diameter / 2. More...

kernelType	m_kernelType
	The kerneltype; displayed as int. More...

core::Img32f	m_kernelImage
	An internal representation of the kernel used. More...

Detailed Description

MeanShiftTracker

The MeanShiftTracker is a tracking component using the MeanShift algorithm.

Given a, preferably formatGrey, so called weightImage, it can find a distribution maxima of the feature tracked in the workingImage. The workingImage must contain weights for each pixel, determing the likelihood of the pixel belonging to the feature being tracked. High values correspond to a high likelihood.

The MeanShiftTracker uses the MeanShift algorithm for tracking blobs in realtime. It is an efficient technique for tracking 2D blobs through an image. Given a, preferably formatGrey, weightImage, it can find a distribution maxima of the feature tracked. The weightImage must contain weights for each pixel, determing the likelihood of the pixel belonging to the blob. High values correspond to a high likelihood. The Meanshift algorithm is resistent to camera movements, partial covering, scaling and noise. The Meanshift algorithm is also useful in image segmentation.

The MeanShift algorithm

As stated above, the MeanShiftTracker uses the MeanShift algorithm to find the new center of the blob being tracked. This is the MeanShift algorithm:

$P_n = \frac{\sum_{A_h}k(A_h-P_o)w(A_h)P_o}{\sum_{A_h} | k(A_h-P_o) w(A_h)|}$

, where $P_n $ is the new center of the blob, $P_o $ is the previous center, where the algorithm starts, $k() $ is the kernel function, $w() $ is the weight function and $A_h $ are all pixels located in a squarish region of the size 2*bandwidth around $P_o $ .

The Bandwidth

The Bandwidth is the only parameter for this algorithm. It is half the length of the squarish mask containing all pixels $A$ . Choosing a correct size for this mask is essential. It should approximatly cover the blob being tracked.

The Kernel

The Kernel determines how much each pixel $A$ contributes to the outcome of the formula. Using different kernels changes the speed and precision of the algorithm. So far, the Epanechnikov kernel, which is used as the default kernel for this algorithm, due to its good performance, and a Gauss kernel have been implemented.

The weightImage

This implementation of the Meanshift algorithm does not work with the original image. Only a one-channel image, having the same size and using the same coordinate system as the original image, is needed. In this image, all pixels, that belong to the blob must have high values.

ToDo

Add other kernels. Add functionality to open any image as kernel image. Performance tweaking.

Member Enumeration Documentation

◆ kernelType

enum icl::cv::MeanShiftTracker::kernelType

An enumeration for the different kernel types.

Enumerator
epanechnikov
gauss

Constructor & Destructor Documentation

◆ MeanShiftTracker()

icl::cv::MeanShiftTracker::MeanShiftTracker	(	kernelType	type,
		int	bandwidth,
		float	stdDev = `1`
	)

Constructor with only the most needed parameters.

The basic constructor with only the most needed parameters.

Parameters

type	The desired kernel type
bandwidth	The desired kernel bandwidth
stdDev	profiles the standard deviation.

Member Function Documentation

◆ applyMeanShiftStep()

utils::Point32f icl::cv::MeanShiftTracker::applyMeanShiftStep	(	const core::Img32f &	image,
		const utils::Point32f &	pos
	)

private

Applies a single step of the mean shift algorithm.

Parameters

image	image with which the mean shift step will be performed
pos	start original position

Returns: new center

◆ generateEpanechnikov()

static core::Img32f icl::cv::MeanShiftTracker::generateEpanechnikov ( int bandwidth )

static

Generates an Epanechnikov Kernel.

Parameters

bandwidth kernel bandwidth

◆ generateGauss()

static core::Img32f icl::cv::MeanShiftTracker::generateGauss	(	int	bandwidth,
		float	stdDev
	)

static

Generates a Gauss Kernel.

Generates the internal m_opKernelImage using the Gauss profile

Parameters

bandwidth	kernel bandwidth
stdDev	profiles the standard deviation.

◆ getBandwidth()

int icl::cv::MeanShiftTracker::getBandwidth ( ) const

inline

Returns current kernel bandwidth.

◆ getKernel()

int icl::cv::MeanShiftTracker::getKernel ( ) const

inline

returns current kernel type

◆ getKernelImage()

const core::Img32f& icl::cv::MeanShiftTracker::getKernelImage ( ) const

inline

returns current kernel image

◆ setKernel()

void icl::cv::MeanShiftTracker::setKernel	(	kernelType	type,
		int	bandwidth,
		float	stdDev = `1`
	)

sets new kernel type (with params)

Parameters

type	kernel type
bandwidth	new bandwidth
stdDev	profiles the standard deviation.

◆ step()

const utils::Point32f icl::cv::MeanShiftTracker::step	(	const core::Img32f &	weigthImage,
		const utils::Point32f &	initialPoint,
		int	maxCycles = `-1`,
		float	convergenceCriterion = `1.0`,
		bool *	converged = `0`
	)

This function returns a new center after the MeanShift algorithm is applied.

Parameters

weigthImage	gray level input image
initialPoint	starting point for meanshift loop
maxCycles	A treshold for the maximum iteration coutn if (-1, 10000 is used as upper limit)
convergenceCriterion	position difference (in pixels) of consecutive steps that is used as threshold for mean shift loop
converged	if a non-NULL pointer is given here, the reference boolean is used to notify the caller whether actually maxCycles iteration steps were used or the convergence criterion was reached

Member Data Documentation

◆ m_bandwidth

int icl::cv::MeanShiftTracker::m_bandwidth

private

The bandwidth of the kernel; equals diameter / 2.

◆ m_kernelImage

core::Img32f icl::cv::MeanShiftTracker::m_kernelImage

private

An internal representation of the kernel used.

An image of the kernel used for the MeanShift. Stored is only the second quadrant of the image, as the values are the same for absolute coordinates.

◆ m_kernelType

kernelType icl::cv::MeanShiftTracker::m_kernelType

private

The kerneltype; displayed as int.

The documentation for this class was generated from the following file:

/Users/alneuman/vm/icl/ICLCV/src/ICLCV/MeanShiftTracker.h

Public Types

Public Member Functions

Static Public Member Functions

Private Member Functions

Private Attributes

Detailed Description

The MeanShift algorithm

The Bandwidth

The Kernel

The weightImage

ToDo

Member Enumeration Documentation

◆ kernelType

Constructor & Destructor Documentation

◆ MeanShiftTracker()

Member Function Documentation

◆ applyMeanShiftStep()

◆ generateEpanechnikov()

◆ generateGauss()

◆ getBandwidth()

◆ getKernel()

◆ getKernelImage()

◆ setKernel()

◆ step()

Member Data Documentation

◆ m_bandwidth

◆ m_kernelImage

◆ m_kernelType