CLASSIFICATION OF ALPINE SKIERS SKILL LEVEL USING SMARTPHONE DATA
MARTINEZ-ALVAREZ, A., SNYDER, C., NEUWIRTH, C., STÖGGL, T. (2020): CLASSIFICATION OF ALPINE SKIERS SKILL LEVEL USING SMARTPHONE DATA In: Book of Abstracts of the 25th Annual Congress of the European College of Sport Science
The use of smartphones as activity trackers over the last few years has increased exponentially. The release of different sensors and apps have amplified the possibilities of tracking and evaluating alpine skiing. However, the wide range of skier skill levels is a factor to take into consideration when comparing metrics between skiers or including gamification into tracking. Consequently, the aim of this project was to develop a skill level classification tool for alpine skiers based on smartphone data.
METHODS:
Data was collected from 31 skiers (14 intermediate, 11 advanced and 6 expert). The skiers were instructed to simply turn on the app and keep it running through the entire session. They had to ski for at least half a day, record a video of them skiing and fill a questionnaire after the skiing day.
The data was collected using a customized app that recorded barometer, GPS, accelerometer and gyroscope signal from the smartphone sensors. The video was used by expert raters to assess the level of the skiers. The questionnaire was used to report if they were skiing alone or in a group and the skiing background.
Shapiro-Wilk test was used to assess normality. To avoid multicollinearity, variables with a correlation higher than 0.8 were removed. The between group comparisons were assessed by means of 1-way ANOVA and Kruskal-Wallis test for normal and non-normally distributed variables respectively. Bonferroni post-hoc pairwise comparisons were used when significant differences were found. To classify the data into the three groups, a decision tree was trained. Significance was assessed at α = 0.05.
RESULTS:
An activity detection algorithm to differentiate between skiing, being on the lift and being stopped was developed. Different thresholds such as instantaneous speed and altitude difference were used to determine the outcome. Based on those results and the raw data 44 metrics were extracted, such as: speed, active time, pause duration, altitude changes, etc. After multicollinearity analysis 30 variables were included for further analysis. Differences between groups were found for maximum speed (p<0.001), average speed (p<0.001), minutes skiing (p=0.001), minutes stopped before the lift (p=0.039), total pause time (p=0.049), number of stops during runs (p=0.011), and the ratio between the active time and the time at the bottom of the lift (p=0.022). The results of the decision tree showed a training accuracy of 80.6% based on two variables: average speed (25 km/h) and total pause during skiing (68’). CONCLUSION: The results of this study highlight differences in some metrics between intermediate, advanced and expert skiers. Although the decision tree classifies properly 80.6%, all the data was used as a training set due to the small sample. We conclude that the results indicate the feasibility of classifying the skiers’ skill based on smartphone data but more data is needed to further improve and validate the classification model.
