AI Case Study
Researchers at Chalmers University of Technology classify pole strokes in skiing using machine learning
Researchers at Chalmers University of Technology have leveraged machine learning to provide better technique analytics for ski athletes. They programmed a system to classify pole strokes by analysing data from poles with power meters using machine learning. Using long short-term memory (LSTM) they achieved 95% accuracy in classification, and 78% accuracy when using data from only two skiers and testing on the third.
"A team of scientists hailing from Chalmers University of Technology in Sweden describes AI that distinguishes among ski techniques from data collected by sensor-packed ski poles. They believe it could help coaches and athletes to analyze things like training load and racing efforts, and to make adjustments accordingly in real time.
“In this project, we have collaborated with [a company] which produces a power meter for cross-country skiing, mounted inside the handle of the pole,” the paper’s coauthors wrote.
As they further explain, ski techniques can be broadly divided into two styles — classical style and freestyle — and various sub-techniques (or gears), each of which is uniquely suited to distinct terrain types and snow conditions. The scientists focused on double poling (which is primarily used on gentle downhill slopes) in the course of their research, but they also considered “gear 2,” an uphill technique; “gear 3,” which is used to transition between uphill and downhill skiing; and “gear 4”, a horizontal move.
Post-training, the researchers evaluated the models on an unseen subset of data. They report that the best-performing one — a long short-term memory (LSTM) network, a type of recurrent neural network that can “remember” values over an arbitrary length of time — correctly classified 95% of strokes. Unsurprisingly, accuracy dropped steeply (to 78%) when the models were trained on data from only two skiers and tested on samples from third.
The researchers note that their technique effectively only records movement from the hands and that it doesn’t include sensors on the body or skis. Still, they believe that, given a larger corpus containing data from both professionals and recreational skiers, the models could achieve even higher classification accuracy.
“To achieve better generalization to individuals not appearing in the training set more data is required, which is ongoing work,” they wrote. “Nevertheless, we reach comparable or better … results by using [AI] models, something which has not been much explored in other studies and has the advantage of not requiring handcrafted features to be passed to the model.”
"Power meters are becoming a widely used tool for measuring training and racing effort in cycling, and are now spreading also to other sports. This means that increasing volumes of data can be collected from athletes, with the aim of helping coaches and athletes analyse and understanding training load, racing efforts, technique etc." (paper)
The system was able to accurately classify 95% of 1,671 individual pole strokes when trained on data from three skiers and 78% when trained on data from two skiers and being tested on data from the third.
"The team sourced a data set provided by Skisens AB, a spinoff company from Chalmers, containing samples from three skiers using sensor-equipped handles in different three locations. They extracted 1,671 individual pole strokes in the course of preprocessing the data, which they fed into three different machine learning models for classification."