Using the Hackel article in /doc, on page "180" is a set of features to be computed for every cell. Disregard the construct they use to build the local tensor and just focus on each feature using the matrix of points for that cell. Store the features inside a vector
PS: Eigen makes it very easy.
Reminder:
Centroid = M.colwise().mean();
Covariance Matrix = (1/nbpoints) ( (M-centroid) * (M-centroid)T )