Gesture Classification using a GMM Front End and Hidden Markov Models

C. Gurrapu and V. Chandran (Australia)


Gesture recognition, Polygon vertices, HMMs, Pose detection, Computer vision


While performing a gesture the human body shape goes through a sequence of changes. We present a system where we use shape description features in the form of polygonal approximation of the outline of the body to describe the in stantaneous shape of the human body. These features are modelled with a GMM, whose parameters in turn form the feature vector of an HMM. The use of this GMM front end allows for the number of shape description features to be time varying. While more general, fixed length, shape de scription features such as Fourier descriptors, eigenvalues etc. exist, we believe the use of features such as polygons or medial axis allows for easier inclusion of context and a pri ori knowledge into the system. Based on this technique, we present a system that can classify 3 different gestures per formed by two different people with near 98% accuracy.

Important Links:

Go Back