Human features extraction by using anatomical and low level image descriptors from whole body images

Interaction experience in multimedia systems can be improved by adding personalization. Current applications for building and animating characters to represent real users are typically based on pose and motion detection. For so doing, computer vision algorithms do not exploit the anatomical characteristics of the human body for improving their classification accuracy. This work presents an strategy that considers age-group, body shape and height estimation by using anatomical low level descriptors. The proposed strategy allows to differentiate children from adults, and under-weighted and normal body shaped from over-weighted individuals, based on a set of features extracted from full body images and a classification process based on Support Vector Machine (SVM). These classification models were evaluated using a 10-fold cross validation, obtaining an area under the ROC curve of 89 % and 92 % respectively for age-group and body shape. On the other hand, the height of a person was computed by using a reference image in a leave-one-out evaluation and, in comparison with the real one, an square error (MSE) of 17cm was obtained.