In this paper we present an approach for capturing 3D body motion and inferring human body posture from detected silhouettes. We show that the integration of two or more silhouettes allows us to perform a 3D body reconstruction while each silhouette can be used for identifying human body postures. The 3D reconstruction is based on the representation of body parts using Generalized Cylinders providing an estimation of the 3D shape of the human body. The 3D shape description is refined by fitting an articulated body model using a particle filter technique. Identifying human body posture from the 2D silhouettes can reduce the complexity of the particle filtering by reducing the search space. We present an appearance-based learning method that uses a shape descriptor of the 2D silhouette for classifying and identifying human posture. The proposed method does not require an articulated body model fitted onto the reconstructed 3D geometry of the human body: It complements the articulated body model since we can define a mapping between the observed shape and the learned descriptions for inferring the articulated body model.