Global Motion Estimation from Pan-Tilt Cameras

Roman Bachmann

Supervisor(s): Helge Rhodin, Pascal Fua

École Polytechnique Federale de Lausanne

Abstract: Coaches in alpine skiing would like to know the speed and useful biomechanical variables at each turn in a run. Existing methods using body-worn sensors are distracting and marker-based manual image annotation for inference is time consuming. We propose a method of estimating an athlete’s global 3D pose using multiple cameras. First, tight estimated bounding boxes of the skier are fed to a 2D pose estimator network. The 3D pose is then reconstructed using a bundle adjustment method. We show results both when using fully calibrated cameras, as well as when estimating the rotation of Pan-Tilt-Zoom cameras. To overcome shortcomings of existing datasets we created a new alpine skiing dataset and trained all methods on it. Our method estimates accurate global 3D poses from images only, providing coaches with an automatic and fast tool to improve an athlete’s performance.
Keywords: 3D Reconstruction, Computer Vision, Image Processing
Full text:
Year: 2019