Inertial sensors based cross-country skiing technique classification and sensors layout optimization

Abstract

Cross-country (XC) skiing has been adopted since 1924 in the Winter Olympics in Chamonix, France. It is a winter sport that traverses the diverse snow-covered terrain. The preferred techniques differ depending on the terrain, and it is known that the result of the competition can be changed even in a few seconds. Therefore, strategic selection of skiing technique according to terrain is considered as an important factor. As a result, the scientific coaching system in XC-skiing has high value and demand for improving sports performance and selection of proper strategic skiing technique. The existing approaches of coaching focused on the methods of teaching through professional experience or through video shooting. Various sensor based analytical methods have been proposed during recent years. Representative sensors include vision sensor, heart rate sensor and foot-pressure sensor. Very recently, due to the development of Micro-Electro-Mechanical System, XC-skiing sports analytics using wearable inertial sensors has been actively conducted, and classification of XC-skiing techniques is regarded as a precedent for strategic selection of skiing techniques.

The main purpose of this study was to automatically classify the XC-skiing techniques using deep learning/machine learning based on the data from wearable inertial sensors and to optimize the sensor layout for future XC skiing analytics. The motion data for training and testing were collected from three XC professional skiers in the snow-covered environment separately. In this study, CNN-LSTM based deep learning model and KNN machine learning model were proposed to classify eight techniques used in classical and skating styles XC-skiing. The proposed deep learning model resulted in high mean classification accuracy of 87.2% and 95.1% for two test datasets using a five-sensor configuration (both hands, both feet, and the pelvis), which is superior to the traditional machine learning method of KNN.

Hierarchical cluster analysis was applied to further find an optimal number of sensors and sensor locations. The optimal sensor layout consists of trunk including head, arms, and legs (three sensors configuration: {body, one arm, one leg}, four sensors configuration: {body, one arm, two legs}, five sensor configurations: {body, two arms,two legs}). Results of usability questionnaire from 6 professional XC-skiers showed that chest and upper leg were the preferred locations for attaching upper body and lower body sensors, respectively.

The findings from this study indicated that the developed CNN-LSTM based unified classification model has the potential to be deployed for real-time classification of major skiing techniques by professional skiers and coaches. In addition, the identified optimal sensor layout may provide promising directions for designing a practical body sensor network for future XC-skiing analytics.