Reconstruction of 3d Human Motion Pose from Uncalibrated Monocular Video Sequences

Abstract. This paper investigates the problem of reconstruction of 3D human motion pose from uncalibrated monocular video sequences. Firstly, we prove that the transform between the change of Z coordinate, (dz), and the corresponding change of scale factor s, (ds), is a linear transformation under perspective camera model. Then, we locate a frame in which the human figure with three linked body segments, the first segment and the last segment are parallel to the image plane. By using these three segments’ coordinates on the image plane and human skeleton proportions constrain, we obtain the precise transform parameter. We establish an equation to obtain the relative depth ∆Z of body segments based on perspective camera model in each frame, the rotation vector of joints is obtained by using IK (inverse-kinematics) based on the obtained relative coordinates of joints of segments in 3D space. Finally, we demonstrate the accuracy, advantages by applying this algorithm to real video sequences.