The aim of this study was to design and validate a sub-technique classification algorithm for everyday cross-country trainings, based on a single sensor worn on the upper back and biomechanical models. The sensor (FieldWiz, Advanced Sport Instruments, Switzerland) recorded GNSS data, acceleration, and angular velocity. Using a customized fusion algorithm and a trunk model, the athlete`s centre of mass kinematics were obtained. Cycles were detected based on maxima in trunk inclination and a Gaussian mixture model was used to assign each cycle to its corresponding sub-technique (Gear 2, 3, 4) based on cycle distance, amount of lateral excursion and trunk inclination periodicity. Gaussian mixture parameters were determined from a separate dataset of short roller skiing trials. The algorithm was validated against video recordings with 5 junior level athletes skating on a 2.4 km lap with roller skis at medium intensity. Turns were removed and uphill and flat sections were selected. 925 sec of data remained and each second was attributed one sub-technique. Gears 2, 3, 4 were skied during 81, 600, 244 sec, respectively. 98.4% of all seconds were correctly classified and misclassifications mainly happened during transitions. This approach of model-based sub-technique classification proved extremely efficient, is fully automatic, and can be used during daily trainings. On-snow validity should be assessed in the future and other sub-techniques (e.g. double poling) could be added.
© Copyright 2019 German Journal of Exercise and Sport Research. Springer. Alle Rechte vorbehalten. / Übersetzt mit https://www.DeepL.com/Translator