Deep learning-based 2D keypoint detection in alpine ski racing – A performance analysis of state-of-the-art algorithms applied to regular skiing and injury situations

• DOI: 10.1016/j.jsampl.2023.100034
• Publication Date: 2023-08-23T16:42:39Z
• Authors: Michael Zwölfer; Dieter Heinrich; Kurt Schindelwig; Bastian Wandt; Helge Rhodin; Jörg Spörri; Werner Nachbauer
• Keywords: ACL injuries; Athletes; Human pose estimation; GV557-1198.995; Sports medicine; Alpine skiing; 610 Medicine & health; 10046 Balgrist University Hospital, Swiss Spinal Cord Injury Center; Deep learning; RC1200-1245; Sports

In this study, we examined the practicability of deep learning-based 2D keypoint detection applied to regular skiing and injury situations (i.e., out-of-balance fall situations) on an alpine ski racing track. We therefore created a skiing- situation-specific dataset (hereinafter called "Injury Ski Dataset"), which state-of-the-art algorithms OpenPose, Mask-R-CNN, AlphaPose DCPose were compared. The performance each detector was evaluated by calculating mean per joint position error (MPJPE) percentage correct keypoints (PCK). Failure cases common patterns further investigated visual analysis. observed best results for skiing, with 81%–92% all detected correctly at MPJPE 9 (2) 14 (3) pixels. situations, self-occlusions rare poses became more likely, similar occlusions due snow spray motion blur. As result, in decreased 68%–80% (PCK), while only 35%–54% correctly, errors 26–36 Among algorithms, most robust achieved results. PCK range manual annotation can be considered low enough biomechanical For should improved. Regarding development learning tool analysis future, propose fine-tune well-performing detector, such as AlphaPose, ski- injury-specific dataset, ours.