3D ball tracker

This page contains information on how to build, set up and run pf3dTracker, the particle-filter-based 3D ball tracker, and pf3dBottomup, the 3D ball detector. The page is in the process of being written, so it's incomplete and the information you find on it might be inaccurate. Should you have any question or complaint, please write Matteo Taiana an email at: mtaiana at isr*ist*utl*pt.

System architecture and behaviour

For the two modules to work well together, they should share the same colour and shape model for the ball, and the same camera model parameters. See below for details on how to write the initialization files accordingly.

Get and build the source code

The source code of tracker and detector is part of the iCub repository, this page explains how to get it, this other one how to build it. Both modules depend on YARP and OpenCV, the tracker depends also on iKin, while the bottom up module depends also on iCubVis. On this page you will encounter the variables $ICUB_ROOT and $ICUB_DIR. $ICUB_ROOT should point to the root of your copy of the iCub repository, while $ICUB_DIR should point to the directory where you build or install the binaries, have a look here for more information. The source code of the tracker is contained in the directory: $ICUB_ROOT/main/src/modules/pf3dTracker, the one for the detector is in: $ICUB_ROOT/main/src/modules/pf3dBottomup. The binaries, after the building process, are stored in the directory: $ICUB_DIR/bin. You should be able to invoke them from any directory.

Set the tracker up

An example configuration comes with the iCub software, so you can test the tracker even without creating the models which are presented hereafter. Beware that the tracker will not work well without customized models.

For the tracker to work properly, you need to create a colour model for the specific ball you want to track. This is done by grabbing images with the camera you want to use, cutting out the parts of the images where the ball is seen and pasting them all together in one file. The background of this image should be white, as white pixels are discarded when building the model histogram. The robustness of the tracker will depend on this model: you should include images in which the ball is seen under different lighting conditions. The more images you cut out, the better. If you can change the colour/brightness parameters of the camera, please do so before creating the colour model and use the same settings every time you use the tracker. Good setting include high saturation and a brightness that never makes parts of the ball appear as white or black.

Two examples of colour template images, for a yellow and a red ball, respectively:

Some images depicting the hand of the iCub robot were included in the template for the red ball, hoping this will improve the tracking when the ball is partially occluded by the hand of the robot.

You need to create a shape model for the specific ball you want to track. This is done using the Matlab script $ICUB_ROOT/src/pf3dTracker/matlab_files/write_initial_ball_points.m. You should set three parameters inside the script: R, R1 and R2. R is the radius of the ball you want to track, in millimetres. R1 and R2 are the radii that are used to project the inner and outer contour (see [[1]] for more details). If you want a precise estimate of the 3D position of the ball, you should set R1 and R2 close to the value of R (e.g. 10% difference). If you want the tracker to be able to withstand high accelerations of the ball, maintaining the number of particles used low, you should increase the difference up to 30% (this is the value I typically use). This script will create a file called something like: initial_ball_points_31mm_30percent.csv.

You need to create a dynamic model for the ball. Basically you have to fill in the dynamic matrix. I use a constant velocity model, with random acceleration. The data for this is stored in: models/motion_model_matrix.csv. I'm not sure that the tracker will work properly with other configurations of the motion model. For the dynamic model it is also quite important the parameter AccelStDev, that is set in the initialization file (see below).

You need to calibrate the camera you use i.e. estimate the intrinsic camera parameters. You can do that using camCalibConf, for example.

You need to customize the file that sets the tracker up on start up. The default initialization file is $ICUB_ROOT/app/pf3dTracker/pf3dTracker.ini. Here is an example:

 ####################################
 #configuration file for pf3dTracker#
 ####################################
 
 
 #############
 #module name#
 #############
 name                        /pf3dTracker
 
 #############################
 #parameters of the algorithm#
 #############################
 nParticles                  900
 #nParticles                 number of particles used
 accelStDev                  30
 #accelStDev                 standard deviation of the acceleration noise
 insideOutsideDiffWeight     1.5
 #insideOutsideDiffWeight    inside-outside difference weight for the likelihood function
 colorTransfPolicy           1
 #colorTransfPolicy          [0=transform the whole image | 1=only transform the pixels you need]
 
 
 #########################
 #port names and function#
 #########################
 inputVideoPort              /pf3dTracker/video:i
 #inputVideoPort             receives images from the grabber or the rectifying program.
 outputVideoPort             /pf3dTracker/video:o
 #outputVideoPort            produces images in which the contour of the estimated ball is highlighted
 outputDataPort              /pf3dTracker/data:o
 #outputDataPort             produces a stream of data in the format: X, Y, Z, likelihood, U, V, seeing_object
 outputParticlePort          /pf3dTracker/particle:o
 #outputParticlePort         produces data for the plotter. it is usually not active for performance reasons.
 outputAttentionPort         /pf3dTracker/attention:o
 #outputAttentionPort        produces data for the attention system, in terms of a peak of saliency.
 
 
 #################################
 #projection model and parameters#
 #################################
 #projectionModel [perspective|equidistance|unified]
 projectionModel             perspective
 
 #iCubLisboaLeftEye_Zoom_Lens_2009_05_19
 w 320
 h 240
 perspectiveFx 445.202
 perspectiveFy 445.664
 perspectiveCx 188.297
 perspectiveCy 138.496
 
 
 #######################
 #tracked object models#
 #######################
 #trackedObjectType [sphere|parallelogram]
 trackedObjectType           sphere
 trackedObjectColorTemplate  models/red_smiley_2009_07_02.bmp
 trackedObjectShapeTemplate  models/initial_ball_points_smiley_31mm_20percent.csv
 
 motionModelMatrix           models/motion_model_matrix.csv
 trackedObjectTemp           current_histogram.csv
 
 
 #######################
 #initialization method#
 #######################
 #initialization method [search|3dEstimate|2dEstimate]
 initializationMethod        3dEstimate
 #initial position [meters]
 initialX                       0
 initialY                       0
 initialZ                       0.5  
 
 
 ####################
 #visualization mode#
 ####################
 #circleVisualizationMode	[0=inner and outer cirlce | 1=one circle with the correct radious]
 #default 0. only applies to the sphere.
 circleVisualizationMode	1
 
 
 #########################
 #attention-related stuff#
 #########################
 #the tracker produces a value of likelihood at each time step.
 #this value can be used to infer if the object it is tracking is the correct one.
 #this is not a very robust way of doing so.
 #if likelihood>this value, then probably I'm tracking the object.
 likelihoodThreshold         0.005
 attentionOutputMax          300
 attentionOutputDecrease     0.99
 
 
 ##########################
 #image saving preferences#
 ##########################
 #save images with OpenCV?
 saveImagesWithOpencv        false
 #always use the trailing slash here.
 saveImagesWithOpencvDir     ./graphical_results/

Run the tracker

The tracker works best if the brightness/colour parameters of the camera are set at the same values they had at the time of the acquisition of the images that form the colour model. To make sure of this, you should note down the values of the camera parameters at the time of the acquisition of such images and set the parameters again to those values before running the tracker (you can use frameGrabberGui for this).

Here is an example of such parameters:

 brightness 0
 sharpness  0.5
 white balance 0.648 0.474
 hue        0.482
 saturation 0.826
 gamma      0.400
 shutter    0.592
 gain       0.305

Run the tracker via the application manager

You need to change the "node" information in the xml file before you run it. This script relies on yarprun. If you do not know what yarprun is, it's probably faster if you try the other method to run the tracker.

  cd $ICUB_ROOT/app/default/scripts
  ./manager.py $ICUB_ROOT/app/demoReach_IIT_ISR/scripts/iit/demoReach_IIT_ISR_NoHand.xml

or

  cd $ICUB_ROOT/app/default/scripts
  ./manager.py $ICUB_ROOT/app/demoReach_IIT_ISR/scripts/isr/demoReach_IIT_ISR_JustTracker.xml

To run the tracker together with the bottom up detection module:

  cd $ICUB_ROOT/app/pf3dTracker/scripts
  manager.py pf3dTrackerWithBottomup.xml

Check the dependencies, run the modules and connect the ports.

Run the tracker invoking commands from the shell

  #run an image rectifier, in case you need it (cameras with a non-negligible distortion)
  camCalib --file iCubLeftEye.ini --name /icub/camcalib/left
  
  #run the tracker itself
  pf3dTracker
  
  #start a viewer
  yarpview /viewer
  
  #connect all the ports
  yarp connect /icub/cam/left /icub/camcalib/left/in
  yarp connect /icub/camcalib/left/out /pf3dTracker/video:i
  yarp connect /pf3dTracker/video:o /viewer

You can specify a context and an initialization file name for the tracker with the options --context and --from. For more information on this matter, please check out these pages on the resource finder: [2], [3].

Theoretical foundations of the tracker

If you want to know more on the theoretical ideas behind the tracker, please have a look at the papers on this page: [4].

Demo videos

If you want to watch videos and evaluate the performance of the tracker, please have a look at this page: [5].

ToDo

Force the tracker to avoid putting particles close to the origin of the reference frame of the camera and behind the camera. This might be causing the transient problem at initialization time.

Make the building of the tracker independent of iKin and thus of IPOPT. Only a part of the code should depend on this.

Translate from Matlab to C++ the piece of code that computes the initial ball points and read the value of the radius from the initialization file

Read the color/illumination parameters from the camera before starting the tracker; set the desired parameters of the camera, then start the tracker; restore the original parameters when quitting

Make the tracker compute the histogram with Gaussian kernels instead of Dirac's

Make the tracker quit gracefully when asked to, instead of requiring multiple ctrl-c's

Make sure the tracker builds and works well on Windows

Document the code with Doxygen

Add the "Expected behaviour" section to the wiki, where the desired behaviour of the tracker is described.

Make the tracker adaptive to different image sizes

Turn the number of particles into a parameter loaded at start time

Get rid of IPP dependency, using OpenCV

Start writing a wiki-based tutorial

(NOT NEEDED ANY MORE) Fix $PF3DTracker/conf/Find_IPP.cmake