Monocular markerless position tracking of elite amateur boxing fighters in real combat situation

Markerless video analysis represents an opportunity for conducting efficient in-situ motion analysis of athletes during competitions. From monocular video data, we propose a robust end-to-end method to automatically capture the 2D trajectory of athletes' position on planar ground, even in highly occluded contexts. A tracking-by-detection algorithm is first applied on a short sequence to build a specific contextual dataset ('self-supervision'). These data are subsequently used to train a specific person detector. Afterwards, body anatomical features in image coordinates are identified using human pose estimator. Athletes position is extracted as the midpoint between the feet and converted to metric units through homography. The accuracy of our monocular algorithm was evaluated by comparison with a position trajectories calculated from markerless reconstruction of 3D poses using 11 accurately synchronized and calibrated cameras as reference. The average error was 0.3 m over about 130,000 frames at 50 fps. The trajectories of the monocular method and the multiple views reference show an average correlation above 0.9. The robustness of the monocular method was tested in real competition of boxing combats for 18 rounds involving 22 elite fighters. These results open perspectives to provide performance indicators such as ring generalship to the coaching staff with minimal setup.