Background
Cross-country skiing (XC) is one of the most demanding endurance sports in the Alpine region. In this study we concentrate on the skating discipline of XC, in which sub-technique choice is particularly variable. Athletes permanently switch between different “sub-techniques” depending on the slope, their speed and the snow conditions. The sub-technique chosen has a strong effect on the athlete’s energy economy, pacing and race performance. For coaches, the analysis of the athlete’s sub-technique choices is an important factor for training and race analysis. Today coaches at Swiss-Ski analyse sub-techniques manually from video, which is tedious and resource-intensive. Sensor-based classification technology exists internationally. However, side-specific differentiation of sub-techniques remains challenging in single-sensor setups.
Aim
The aim of the study is to develop a sensor-based technology that meets elite XC racing demands and will be used by Swiss-Ski in the future. The technology must reach 4 goals simultaneously: i) a single light-weight sensor, ii) high accuracy identification of sub-techniques and sub-technique-transitions, iii) low number of training data needed (allowing in the future for individualized models) and iv) usage of a lightweight explainable AI algorithm.
Methods
Thirteen sub-elite Swiss-Ski athletes performed structured roller-ski tests on a large motorised treadmill at the National Performance Centre Davos. An inertial measurement unit (IMU) recorded acceleration and rotation of the upper trunk at 200 Hz, while a time-synchronised video allowed to identify ground truth data by visual inspection. Each movement cycle was automatically detected from the sideways rotation of the trunk and described by 85 interpretable features capturing amplitude, shape, symmetry and timing. A random forest classifier with 400 decision trees was trained on roughly 12,000 cycles to distinguish five side-specific gears (G2 left/right, G3, G4 left/right). Two competition-like validation protocols, performed without technique instructions, served as an independent test set. Random forest classification was chosen because it requires little training data, has low resource demands, and produces explainable results.
Results
On 3,551 independent validation cycles, the model reached 98.1% overall accuracy. A leave-one-athlete-out test, in which the model was repeatedly retrained while excluding one athlete, confirmed that the system generalises to people it has never seen (97.7 ± 4.1 %). Sub-technique transitions were detected with an F1 score of 0.93 and a mean timing error of only 0.06 seconds within a ±2-second tolerance window. The most informative features described how the trunk leans sideways and rocks forwards and backwards during each cycle — rotational rather than linear motion. This matches what coaches observe by eye, and helps to explain why a single back-mounted sensor is enough.
Discussion and outlook
The result shows that high recognition accuracy does not require heavy deep-learning models or multiple sensors. A compact, interpretable random forest classifier, fed with biomechanically meaningful trunk-motion features, is sufficient — and small enough to run on a watch or sensor in the future. The current validation took place on a treadmill, which keeps conditions clean. A follow-up field study, currently underway with Swiss-Ski, examines how the system performs on real snow with varying terrain. If it transfers well, coaches and athletes could move from labour-intensive manual video tagging to continuous, objective technique analysis — a practical contribution from Graubünden to Alpine winter sport.