Viewpoint-dependent 3D human body posing for sports legacy recovery from images and video

In this paper we present a method for 3D human body pose reconstruction from images and video, in the context of sports legacy recovery. The video and image legacy content can include camera motion, several players, considerable partial occlusions, motion blur and image noise, recorded with non-calibrated cameras, which increases even more the difficulty of solving the problem of 3D reconstruction from 2D data. Therefore, we propose a semi-automatic approach in which a set of 2D key-points are manually marked in key-frames and then an automatic process estimates the camera calibration parameters, the positions and poses of the players and their body part dimensions. In-between frames are automatically estimated taking into account constraints related to human kinematics and collisions with the environment. Experimental results show that this approach obtains reconstructions that can help to analyze playing techniques and the evolution of sports through time.