I'm about to walk out the door so I can't look up papers, but I'm going to guess that it's based as much on dynamics as actual proportions of limbs. I would estimate the number of features to be on the scale of 2000-2010 era facial recognition. Dynamics would increase the number of dimensions, making it possible to separate two people with nearly identical features.
However, I do agree that inferring the proportions from video imagery at various angles and with various items of clothing is going to be challenging due to systemic distortion, e.g. inferring incorrect angle due to clothing.
However, I do agree that inferring the proportions from video imagery at various angles and with various items of clothing is going to be challenging due to systemic distortion, e.g. inferring incorrect angle due to clothing.