The authors focused on identifying the necessary elements to use an HMM system regardless of the sensor device being used. It would work as long as it has information about the 3 axis of motion. In a hidden Markov model, a sequence is modeled as an output generated by a stochastic process progressing through discrete time steps, where a symbol from the alphabet is outputted at each step. Only the sequence of emitted symbols is observed. HMM requires a 1:1 ratio of state to alphabet. They carved their space into subcubes, where they got alphabet sizes of 27, 64, and 125. 27 was the chosen size since recognition time decreased and were able to achieve recognition of 800 gestures in a second. 250 samples in a training set is good for detection results.
In testing, the user would press a button, perform the gesture, and let go. less than 27 was considered short and greater than 27 was considered long. They determined that left and right hand data sets were different enough to throw off the results.
Discussion
I still don't know much about HMM. It would take a while to fully understand it. That's the reason why I didn't use them in my robotics project.
------------------------------------------------------
Anthony Whitehead, Kaitlyn Fox. Device Agnostic 3D Gesture Recognition using Hidden Markov Models. GDC Canada 2009.
No comments:
Post a Comment