INTRODUCTION: In both classical and skating styles, choosing the right sub-techniques is key to reducing energy use and enhancing performance. Tracking these choices during training and competition provides insights into race strategies and highlights areas for physical improvement. While previous studies have proposed automatic detection systems for sub-technique identification, none have utilized commercially available acquisition systems. This study aims to develop a machine learning algorithm capable of accurately identifying techniques and sub-techniques in XC skiing, leveraging existing technology. METHODS: Raw data were acquired at a rate of 100 Hz from IMU sensors fitted into the handles of each of the two poles during cross-country skiing on a snow track, using both classical and skating techniques. Eight athletes participated in the data collection. A researcher labeled different parts of the data as specific sub-techniques, based on audio and video recordings taken during the skiing session. IMU data were used to estimate the inclination of the poles, with these estimates validated against 3D motion capture. The estimated angles were used to segment the recordings by leveraging peak angles of the poles. For each segmented window, a vector of 37 features was computed: 18 features per pole and an additional feature considering the phase delay between the poles. This vector was then used to train a multiclass Support Vector Machine (SVM) capable of classifying nine distinct skiing styles. To prevent overfitting, a stratified k-fold validation (k=5) approach was applied. RESULTS/DISCUSSION: Data from all subjects were used to train the SVM algorithm, resulting in a subject-independent classifier. Classification performance was evaluated using standard metrics: accuracy (87%), precision (88%), recall (87%), and F1-score (87%). Additionally, a confusion matrix was employed to assess where errors in the classification model were made. The confusion matrix highlights that most classification errors occur between "double poling" (DP) and "double poling with kick" (DK), as well as between "G4_Rigth" with "G4_Left" (G4_L). CONCLUSION: This method, using only data from sensors inside the poles, is the first known to be capable of automatically classifying skiers' techniques—both classical and skating—and also differentiating 'strong' and 'weak' side movements in skating.
© Copyright 2025 10th International Congress on Science and Skiing, January 28 - February 1, 2025, Val di Fiemme, Italy. All rights reserved.