A question about Hidden Markov Model Classifier

[Omar Amin]

Dear Cesar,

I hope this message finds you well, I've built a hidden markov model classifier for 3d gesture recognition using microsoft Kinect and i'm able to classify 15 3d gestures that involve one handed gestures and two handed gestures. with accuracy around 95%

I've two problems that i hope i can find your advice or just simple guidance to solve them.

1- First problem is with scalability, how to extend my solution to classify more than 15 gesture, for around 200 gestures.

I'm asking for your intuition, is that possible if all gestures to be classified contains different hand movement, and i'm building the feature vector from right hand and left hand coordinates after subtracting a stable joint to compensate for the person position in front of the Kinect.

2- Second problem is, when some hidden markov models are dominant after the baum welch learning, as some gestures are misclassified a lot to a specific gesture, why this happen, why a certain HMM corresponding to a certain gesture will always be dominant and will always be in the first three most probable signs, regardless of whether the gesture being classified is like that specific gesture or not.

My Trials that i've tried up till now are:

- I'm using one hidden markov model per gesture,

- I'm using multivariatenormaldistribution as my distribution also i've tried mixture of gaussians.

- i've tried tying both hands movement with the same multivariategaussian with 6 dimensions (right hand x, right hand y, rz,lx,ly,lz)

- also i've tried building two classifiers for each hand, and sum the log probability if the gesture is two handed.

[ThorPedo]

Hi Omar

how are you doing with your application.  Are you still working on it?  how has your experience with the Accord.net library been.  Have you used SVM part of it for classification.  Are you interested in exchanging information on projects and experience with using Machine Learning in applications?  I am a developer and working in field of Computer Vision and related topics.  Now trying to implement a visual classifier.  plesae contact me if you are interested.

Thor