Inference of human postures by classification of 3D human body shape

In this paper we describe an approach for inferring the body posture using a 3D visual-hull constructed from a set of silhouettes. We introduce an appearance-based, view-independent, 3D shape description for classifying and identifying human posture using a support vector machine. The proposed global shape description is invariant to rotation, scale and translation and varies continuously with 3D shape variations. This shape representation is used for training a support vector machine allowing the characterization of human body postures from the computed visual hull. The main advantage of the shape description is its ability to capture human shape variation allowing the identification of body postures across multiple people. The proposed method is illustrated on a set of video streams of body postures captured by four synchronous cameras.