My interest in gesture recognition was mainly an offshoot of my object detection project. My interaction with Prof. Mathias Kolsch and his thesis has contributed a great deal in my interest in improving feature based gesture recognition.
I would also like to thank Prof. Matthew Turk and the whole team at the 4 eyes lab at the University of California at Santa Barbara for the HandVu libraries. The results of my work, with the significant additions and improvements particularly in the detection stage has been published here. A free copy can be obtained from my publications page.
Since this project was mainly with the intention of making a gesture recognition interface a viable alternative for consumer electronics device, many of the routines of the original code has been removed. Particularly, the Flock of features has been extensively cut down. The thesis by Kolsch was mainly intended as a wearable computer where the background varies much more dynamically. Hence the detection stage is much more integral to his work.
However, for the problem of gesture recognition for stationary devices, it is computationally inefficient to use the detection stage on each and every frame of the camera. So an initial background/foreground stage is used to reduce the time of detection. This along with the optical flow creates a very specific area of interest within which the detection area can be confined. And also the detection is not done on each and every frame. It is done only if the movement is "limited" and hence is probably tracking the wrong object.