Augmented poselets for human body pose inference by a probabilistic graphical model

Human body pose estimation is a challenging task which, depending on the context of application, degree of accuracy and availability of image frames, is being faced with different state-of-the-art approaches. In this paper we propose a part-based detection approach combined with a probabilistic graphical model framework for structural constraints on monocular single images, which offers several benefits: human body joints localization inference (rather than a holistic body detection), low computational cost, and robustness against unknown poses as long as antropomorphic constraints are preserved. These outcomes make this approach feasible for applications related to portable devices or multimedia applications which need to be aware of the presence of people in real time at a low cost, and can take advantage of the knowledge about body poses. The presented approach is built by taking into account the existing "Poselets" architecture and one of its foundations, the "H3D" dataset. On top of this, we "augment" the prior knowledge about human body structure and parts appearance in order to learn spatial probability distributions on body natural constraints, which will be used afterwards by the probabilistic graphical model.