Research Papers

Qi Gan, Sao Mai Nguyen, Stephan Clémençon, Mounim A. El-Yacoubi, Ons Jelassi

Athlete’s pose acquisition and analysis is promising toprovide coaches with details of athletes performance and thus help to improveathletes’ performances with more detailed supervision from coaches. Comparedwith traditional ways of acquiring an athlete's gesture, such as using wearablesensors, computer vision technology has advantages of low-cost, high-efficientand non-intrusive. This paper aims to bridge these two fields, byreconstructing athletes’ trajectory using monocular (i.e. single-camera-shot)videos. Under a few assumptions that are applicable to most of the sports ofathletics, we proposed a method combining computer vision techniques andphysics laws to reconstruct athletes’ trajectories from monocular videos. Themethod first estimates 3D pose of athletes from video inputs, then performskinematic analysis on estimated poses to reconstruct the trajectories ofathletes. We tested this algorithm on videos from the triple jump finals of the2016 Olympics in Rio de Janeiro. We achieved a best performance with 9.1% meanaverage error when using ground-truth foot-ground contact signal and 21.4% meanaverage error when using predicted foot-ground contact signal.