jafar::rtslam::ActiveSearchGrid Class Reference

Active search tesselation grid. More...

---

Detailed Description

Active search tesselation grid.

Author:

jsola

This class implements a tesselation grid for achieving active search behavior in landmark initialization.

The grid defines a set of cells in the image. The idea is to count the number of projected landmarks per grid cell, and use one randomly chosen cell that is empty for feature detection and landmark initialization. This guarantees that the system will automatically populate all the regions of the image.

The feature density can be controlled by adjusting the grid's number of cells. Typically, use grids of 5x5 to 9x9 cells.

This class implements a few interesting features:

• The grid can be randomly re-positioned at each frame to avoid dead zones at the cell edges.
• Only the inner cells are activated for feature detection to avoid reaching the image edges.
• The region of interest (ROI) associated with a particular cell is shrinked with a parametrizable margin to guarantee a minimum separation between existing and new features.

The blue and green grids in the figure below represent the grid at two different offsets, corresponding to two different frames.

The tesselation grid used for active feature detection and initialization

This second figure shows a typical situation that we use to explain the basic mechanism.

A typical configuration of the tesselation grid

Observe the figure and use the following facts as an operation guide:

• The grid is offset by a fraction of a cell size.
  • use renew() at each frame to clear the grid and set a new offset.
• Projected landmarks are represented by red dots.
  • After projection, use addObs() to add a new dot to the grid.
• Cells with projected landmarks inside are 'occupied'.
• Only the inner cells (thick blue rectangle) are considered for Region of Interest (ROI) extraction.
• One cell is chosen randomly among those that are empty.
• The ROI is smaller than the cell to guarantee a minimum feature separation.
  • Use the optional 'separation' parameter at construction time to control this separation.
  • Use getRoi() to obtain an empty ROI for initialization.
• A new feature is to be be searched inside this ROI.
• If you need to search more than one feature per frame, proceed like this:
  • At successful detection, add the detected pixel with addObs().
  • Call getRoi() again.
  • Repeat these two steps for each feature to be searched.

We include here a schematic active-search pseudo-code algorithm to illustrate its operation:

 // Init necessary objects
 ActiveSearch activeSearch;
 ActiveSearchGrid grid(640, 480, 4, 4, 10); // Construction with 10 pixels separation.

 // We start projecting everybody
 for (obs = begin(); obs != end(); obs++)   // loop observations
 {
   obs->project();
   if (obs->isVisible())
     grid.addObs(obs->expectation.x());   // add only visible landmarks
 }

 // Then we process the selected observations
 activeSearch.selectObs();                  // select interesting features
 for (activeSearch.selectedObs);            // loop selected obs
   obs.process();                           // process observation

 // Now we go to initialization
 grid.getRoi(roi);                          // roi is now region of interest
 if (detectFeature(roi))                    // detect inside ROI
   initLandmark();                          // initialize only if successful detection

Definition at line 109 of file activeSearch.hpp.

#include <activeSearch.hpp>

List of all members.

Public Types
typedef std::vector < std::vector < observation_ptr_t > >	Grid
Public Member Functions
	ActiveSearchGrid (const int &imgSize_h, const int &imgSize_v, const int &nCells_h, const int &nCells_v, const int &margin=0, const int &separation=0)
	Constructor.
void	clear ()
	Clear grid.
void	renew ()
	Clear grid and position it at a new random location.
void	addObs (const vec2 &pix, observation_ptr_t obsPtr)
	Add a projected pixel to the grid.
bool	getRoi (image::ConvexRoi &roi)
	Get ROI of a random empty cell.
void	setFailed (const image::ConvexRoi &roi)
	Call this after getRoi if no point was found in the roi in order to avoid searching again in it.
Public Attributes
Grid	projections
Private Member Functions
template<class Vec2 >
veci2	pix2cell (const Vec2 &pix)
	Get cell corresponding to pixel.
veci2	cellOrigin (const veci2 &cell)
	Get cell origin (exact pixel)
vec2	cellCenter (const veci2 &cell)
	Get cell center (can be decimal if size of cell is an odd number of pixels)
bool	pickEmptyCell (veci2 &cell)
	Get one random empty cell.
void	cell2roi (const veci2 &cell, image::ConvexRoi &roi)
	Get the region of interest, reduced by a margin.
Private Attributes
veci2	imgSize
veci2	gridSize
veci2	cellSize
veci2	offset
mati	projectionsCount
mati	emptyCellsTile_tmp
int	separation
int	margin
int	cells_h
Friends
std::ostream &	operator<< (std::ostream &s, ActiveSearchGrid const &grid)

---

Constructor & Destructor Documentation

jafar::rtslam::ActiveSearchGrid::ActiveSearchGrid	(	const int &	imgSize_h,
		const int &	imgSize_v,
		const int &	nCells_h,
		const int &	nCells_v,
		const int &	margin = 0,
		const int &	separation = 0
	)

Constructor.

Parameters:

imgSize_h	horizontal image size, in pixels.
imgSize_v	vertical image size.
nCells_h	horizontal number of cells per image width.
nCells_v	vertical number of cells per image height.
separation	minimum separation between existing and new points.

---

Member Function Documentation

void jafar::rtslam::ActiveSearchGrid::addObs	(	const vec2 &	pix,
		observation_ptr_t	obsPtr
	)

Add a projected pixel to the grid.

Parameters:

pix	the pixel to add.

void jafar::rtslam::ActiveSearchGrid::clear

(

)

Clear grid.

Sets all cell counters to zero.

bool jafar::rtslam::ActiveSearchGrid::getRoi

(

image::ConvexRoi &

roi

)

Get ROI of a random empty cell.

Parameters:

roi	the resulting ROI

Returns:

true if ROI exists.

void jafar::rtslam::ActiveSearchGrid::renew

(

)

Clear grid and position it at a new random location.

Sets all cell counters to zero and sets a new random grid position.

void jafar::rtslam::ActiveSearchGrid::setFailed

(

const image::ConvexRoi &

roi

)

Call this after getRoi if no point was found in the roi in order to avoid searching again in it.

Parameters:

roi	the ROI where nothing was found

---

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

• modules/rtslam/include/rtslam/activeSearch.hpp