step 1:
You need to collect a bunch of negative images. Images that do not have the object that you are trying to track. You need to create a .dat file; a file that is just a list of all the images in a directory.

step 2:
There are a couple of techniques for step two. Opencv can generate positive samples from one image of the object you want to track, this is not the best technique but does work.
when you do this you get a bunch of images that look something like this.


The other approach is to have a series of positive images that are cropped. I think that this is the more accurate approach. I needed figure out how the openCv executable was building these samples. After figuring this out I build an app that loads a series of images that you can click all the objects in the image. You can also use a video camera to add more images and save out images. Here is a demo of the app working.

step 3:
Now you just need to train. Running the opencv haar training executable you will generate an .xml file to load in a haarTracking app. I am still learning all the settings for this step. The perfect tweak of the setting will give you the best result for what you want to track.

Here is what I have been able to get so far. The results are not to bad but I need to learn some more. Any help would be great.

ps: the gui is a new addon im working on download soon, or find it on the OF forum.