Copyright (C) 2011-2012 Brandon L. Reiss
[email protected]

I. INTRODUCTION
This code provides an implementation for a CamShift-based color tracker that
uses an online updating scheme to update the histogram describing the tracked
image patch. The technique takes subdivisions of a predefined size around the
tracker Region of Interest (ROI) output and ranks them based on similarity to
the histogram describing the tracked object color. Patches are then weighted
into the current histogram based on the update weight so that the tracker may
adapt to slight changes in illumination or other appearance changes caused by
the object in motion.

The results from this tracker will vary greatly depending upon:
 - The environment's lighting conditions
 - The quality of the video capture device
 - The framerate of the video capture device
 - Similarity between the tracked object and the background
 - Tuning of the supplied parameters visible as sliders in the tracker GUI

For more information regarding an application of this tracker, see how it was
used as a hand tracker to drive a digital puppet. The writeup of that project
is available at https://files.nyu.edu/na712/public/mocap/demos.html.

Contact the author at [email protected]. Visit http://brandonreiss.com
to learn more about current projects.

II. USAGE

Run the application making sure that you have a capture device connected.
When the demo begins, you will see the tracker GUI along with various windows
displaying live updates of the video stream.

To initialize the tracker, left click and drag the mouse around a ROI in the
main video window. The view will update to indicate the selection, and upon
releasing the mouse it will compute an initial histogram and begin tracking the
colored patch. The ellipse shape is the actual tracker output whereas the box
shows the search window used to initialize the tracker in the next frame. The
tracker uses a simple first-order filter to predict the velocity of the tracked
object and update the track window accordingly.

While tracking, updates to the histogram appear as the colored bars comprising
the histogram change height. Try adjusting the parameters and viewing the
subsequent change in the grayscale backprojection output. This will indicate
how much of the global image region is similar to the histogram being tracked.
To get good results, adjust the parameters so that the tracked object is as
unique as possible in the image. This means that the tracked object has the
most significant blob of white pixels in the backproject view while all other
image regions are dark.

III. COMPILATION

Download and extract the source code to your preferred path SOURCE_ROOT. To
create build materials for your platform, use the following commands
  $ cd SOURCE_ROOT
  $ mkdir build
  $ cd build
  $ cmake ../
where SOURCE_ROOT is the actual path where the source package was extracted. If
CMake gives no errors, then you will have a full set of makefiles, Visual
Studio projects, or other files associated with your platform located in the
build subdirectory of SOURCE_ROOT. See '$ cmake --help' for more information on
supported compilers.

On Windows platforms, you will need to copy manually the Intel Thread Building
Blocks (TBB) DLLs manually into the build folder in order to run the
application.

This application has not been built on Linux at any time recently, although
much of the code should be portable. I'd say give it a shot!