The clusters determined by the above clustering
approach is used to initialize
the feature PDFs for each sub-class of the MR-HMM.
Recall that the MR-HMM needs to estimate the feature PDF
for each combination of segment size and sub-class.
To initialize these PDF estimates, we gather
the original time-series segments from each combination of
cluster index and segment size, and compute
the respective features.
The PDF of the feature collections were then estimated
using a Gaussian mixture.
There is a different PDF estimate for each combination
of cluster (sub-class) index and segment size.
Software for this purpose is given in Section 14.4.2