Predicting Cross-Country Skiing Techniques Using Machine Learning

Abstract

Power meters are vital for a cyclist during training and competitions as a tool for evaluation and improvement. Expanding their application to other sports would help athletes train better during workouts and maximize performance. In crosscountry skiing, power comes from the poles and highly depends on skiing technique used. Thus, in order to predict power, it is important to predict which technique is being used. This thesis focuses on evaluating two machine learning methods to predict techniques in cross-country skiing. Data is collected in collaboration with Skisens AB who provided the sensors and Ulricehamn’s skidgymnasium who helped in collecting data and experimental planning. Data was collected on a treadmill and on roller skis in an outdoor setting to get a balanced set of data and evaluations. Random forests and LSTM networks were selected as the two methods for evaluation. 10 fold cross validation was performed on each model after hyperparameter tuning and the overall accuracy, balanced accuracy and MCC score were recorded. Random forests with a reduced feature set achieved an overall accuracy of 74.4%, while the accuracy of treadmill data and outdoor data was 89.8% and 85.8% respectively. The LSTM model achieved an overall accuracy of 86.2%, while the accuracy of treadmill data and outdoor data was 86.5% and 84.1% respectively.