In this algorithm an eigenspace is formed for every gesture in the vocabulary. By projecting the samples of a gesture into the corresponding eigenspace the manifolds are formed. We segment the manifolds by an unsupervised clustering algorithm. Using vector quantisation the manifolds of the gestures are clustered into an equal number of clusters in each subspace. Each cluster is approximated by a gaussian distribution. Therefore, the set of clusters in an eigenspace is approximated by a sequence of gaussian distributions representing the spatial and temporal variations of hand in the corresponding gesture. We call this sequence a HyperClass.
 

A HyperClass of a sequence of gaussian distributions is treated as a graph in the n-dimensional subspace. Each gaussian distribution in the HyperClass is treated as a vertex in the graph. Therefore the order of the graph is the number of gaussian distributions in the HyperClass.
An unknown gesture is projected into the eigenspaces. The trajectory of the unknown gesture is modeled by a graph in each eigenspace.

The graph of the unknown gesture and the graph of a HyperClass form a bipartite graph in each subspace. We have developed a Graph-Matching technique based on the gaussian probabilities to find the best match between the set of bipartites formed in the subspaces. In this algorithm the graph of the unknown gesture is temporally matched with the graph of the HyperClass in each subspace.