Identifying cross-country skiing techniques using machine learning

In skating-style cross-country skiing, athletes constantly switch between different sub-techniques, and this choice plays a key role in determining their performance. Until now, coaches have had to spend a great deal of time evaluating these techniques using video. In collaboration with Swiss-Ski and the EHSM Magglingen, the University of Applied Sciences of the Grisons has developed a method that automatically and accurately identifies the sub-techniques – down to the specific side – using a single sensor on the back and an interpretable machine-learning model.

In addition, the following individuals were involved in the project:

• Tim Braillard (EHSM)
• Thomas Steiner (EHSM)
• Matheo Käch (Swiss-Ski)

Skating-style cross-country skiing is one of the most demanding endurance disciplines in the Alpine region. Athletes constantly switch between different sub-techniques depending on the incline, speed, and snow conditions. Which sub-technique is used and when has a significant impact on energy efficiency, race pacing, and thus performance.

For coaches, the choice of sub-technique is therefore a crucial part of training and race analysis. At Swiss-Ski, this has so far been done manually using video recordings: accurate, but time-consuming and limited to individual filmed sections of the course.

Wearable sensors could provide this analysis continuously and objectively. However, existing single-sensor systems aggregate the asymmetrical sub-techniques, thereby losing the information – crucial for coaching – regarding which side the skier is leaning toward. This project aims to close that gap.

The goal of the project is to develop sensor-based technology for elite cross-country skiing that can be used by Swiss-Ski in the future. It must meet four requirements simultaneously: a single sensor, highly accurate detection, minimal need for training data, and an explainable, lightweight AI model.

Advantages of sensor-based sub-technique detection:

A single, lightweight sensor

• Only one sensor on the upper back, without athlete-specific calibration
• Data collection over the entire duration of training or a race, rather than just on filmed segments

High accuracy

• Side-specific differentiation of five sub-techniques (G2 left/right, G3, G4 left/right)
• Detection of technique transitions with precise temporal alignment
• Predefined accuracy target of 95% for each of the five classes

Low requirement for training data

• Classification method that requires only a small amount of training data
• Transferability to individuals not included in the training dataset

Explainable and resource-efficient AI

• A transparent model (Random Forest) instead of computationally intensive deep learning
• Model small enough for later operation directly on a sensor or sports watch

Long-term goals are:

• Validation on snow and in varied terrain (ongoing field study with Swiss-Ski)
• Replacing manual video analysis with a continuous, objective technical analysis
• Analysis directly on the sensor or sports watch
• Analysis of side-to-side differences as a new coaching tool

The technology analyzes upper-body movement cycle by cycle. The processing chain consists of four steps:

• A single IMU (ADMOS) on the upper back measures acceleration and angular velocity along three axes each
• The raw signals are smoothed and broken down into individual movement cycles
• For each cycle, interpretable features regarding the amplitude, shape, symmetry, and trajectory of the movement are calculated
• A random forest assigns each cycle to one of five side-specific subtechniques; temporal smoothing yields the output subtechnique and identifies changes as technique transitions

The data set is derived from a study involving 13 elite athletes from Swiss-Ski at the National Performance Center in Davos. On a motorized roller-ski treadmill, they completed a protocol with prescribed techniques as well as two competition-simulated protocols without technique specifications, which served as an independent test dataset. Synchronized video recordings provided the reference.

On independent validation data, the model achieved an overall accuracy of 98.1%; all five sub-techniques exceeded the predefined target of 95%. A «leave-one-athlete-out» test confirmed the model’s generalizability to unknown athletes (97.7 ± 4.1%). Technique transitions were detected with an F1 score of 0.93.

The feature weighting shows that upper-body rotation is the primary factor driving the classification – that is, the movement patterns that coaches also assess visually. Since the results come from a controlled treadmill setting, a field study conducted with Swiss-Ski is currently testing the method’s applicability on snow and variable terrain.

Participants

The project was carried out by the Institute for Data Analysis, Artificial Intelligence, Visualization, and Simulation (DAViS) in collaboration with Swiss-Ski .