Hi dear readers, I've worked on a simple RGB recognition algorithm which is nicely accurate.

Let me show you a little demonstration :

RGB Wheel Tracker from didier.brun on Vimeo.


The algorithm is based on a paper I've read a few years a go, but I can not find the link. The idea is ready simple and require a simple pass on the image.

The first phase is to compare the proportional amount of red, green and blue from 3 relative points, and calculate a score for each pixel.

The second phase is then to blur the result, apply a threshold on the image.

The last phase is to find the blob with the biggest size.

This algorithm is applied for each orientation pattern according to this logic :

  • Iterate the marker searching on each orientation, one by frame rendered (top red, top blue, top green, top red, top blue, ...)
  • If a marker is found, keep the searching on the corresponding orientation only
  • When the marker is lost, iterate on the 3 orientation (point 1) until a marker is found

A sample image :

The algorithm

The result illustrated with "hot" zone :


The PDF marker :  http://www.bytearray.org/wp-content/projects/wheeltracker/rgb_wheel_model.pdf

Live demo : http://www.bytearray.org/wp-content/projects/wheeltracker/wheel_track_demo.swf

Sources : http://www.bytearray.org/wp-content/projects/wheeltracker/sources.zip