Aim To develop an automated markerless 3D motion analysis system for treadmill skiing. Methods A markerless 3D motion analysis was developed using DeepLabCut1 . The 3D model was devel oped using videos from two cameras (LILIN, Taiwan, New Taipei City), that were positioned to the right side and right obliquely from the treadmill. 2D models were trained for both cameras and then calibrated into a 3D model using a calibration frame. The models were trained using 500 images from 7 female and 13 male skiers skiing with V2 skating technique to detect key body landmarks on the skier`s right side, ski pole and roller ski. The analysis was integrated with Coachtech online measurement and feedback system2, and motion analysis results can be combined with other variables computed in Coachtech. Joint an gles obtained with the 3D model were compared to a motion capture system (Vicon, Oxford, UK) consisting of eight cameras and using passive reflective markers on the skier. Results The developed 3D model can detect knee joint angle with root mean square error (RMSE) of 5.4°, ankle joint angle of 7.0° and hip joint angle of 7.8°. Practical applications The automated markerless 3D motion analysis system calculates joint angles using detected points and shows results in Coachtech with only a small delay (< 3 minutes). Thus the system may have several applications for athletes and coaches, in development and monitoring skiing technique. The system is applicable also for other skiing techniques and the accuracy will im prove with additional training data.
© Copyright 2022 Book of Abstracts. 4th Nordic Winter Sports Conference "Performance in snow sports: Translating science into practice" 3rd-4th October, 2022, Mid Sweden University, Campus Östersund, Sweden. Veröffentlicht von Swedish Winter Sports Research Centre at Mid Sweden University. Alle Rechte vorbehalten.