Acquiring human skeleton proportions from monocular images without posture estimation

In this paper we propose a method for estimating human skeleton proportions automatically from two-dimensional (2D) joint locations extracted from a monocular video. Unlike many other methods where the three-dimensional (3D) human postures are pre-known or posture estimations are required, the proposed method does not require correct posture recoveries for the purpose of acquiring the human skeleton model. With a calibrated camera, the proposed system is able to determining the depth of a reference joint for each frame. A human skeleton model is then constructed progressively from that joint and is able to be reprojected to the extracted 2D joint locations under the calibrated camera. The proposed system has been applied to both synthetic and real data and produced highly satisfactory results.