The design of computational methods to recognize human motions is among the most promising research activities in Ambient Intelligence. Accepted solutions use acceleration data provided by wearable sensors. To design general procedures for motion modeling and recognition, this article adopts Gaussian Mixture Modeling and Regression to build computational models of human motion learned from human examples that allow for an easy run-time classification. The main contributions are: (i) an optimized selection of the proper number of Gaussians for building motion models, which is usually assumed to be a priori known; (ii) a comparison between models built by keeping the acceleration axes independent (i.e., 6x2D approach) and models taking axes correlation into account (i.e., referred to as 2 x 4D approach).