Semantic parametric body shape estimation from noisy depth sequences

The paper proposes a complete framework for tracking and modeling articulated human bodies from sequences of range maps acquired from off-the-shelf depth cameras. In particular, we propose an original approach for fitting a pre-defined parametric shape model to depth data by exploiting the 3D body pose tracked through a sequence of range maps. To this goal, we make use of multiple types of constraints and cues embedded into a unique cost function, which is then efficiently minimized. Our framework is able to yield compact semantic tags associated to the estimated body shape by leveraging on semantic body modeling from MakeHuman and L1 relaxation, and relies on the tools and algorithms provided by the open source Point Cloud Library (PCL), representing a good integration of the functionalities available therein.