Automated Classification Of Figure Skating Jumps Using Wearable Sensors

Wearable sensors have the potential to improve training efficiency in the artistic and athletic sport of figure skating. While previous literature estimates that skaters perform 50-100 jumps/day, there is no scientific evidence to support that. Documenting on-ice training load (i.e. number of jumps performed) may be the first step towards creating more efficient training plans, which could reduce overuse injury incidence and improve performance. Data from wearable inertial measurement units (IMUs) can be used to identify when a skater has performed a jump using an automated algorithm. While simply counting jumps would represent progress in monitoring training load, documenting how many of each of the 6 types of multi-revolution jumps were performed could provide athletes and coaches greater perspective on training. PURPOSE: The purpose of this study was to explore whether data from IMUs worn on the ankle and/or waist could be used to differentiate between the 6 jump types. METHODS: Data were recorded from IMUs as 58 competitive figure skaters performed 705 double and triple revolution jumps. The IMUs were worn on the waist and the ankle of the landing leg. Synchronized high speed videos were used to identify jump type and moment of take-off. A random forest machine learning algorithm was trained on processed gyroscope and accelerometer data, and accuracy was assessed using 5-fold cross validation. Due to observed similarities between jump types, 2 classifiers were tested - one to differentiate between 4 categories of jumps and the other to differentiate between all 6 possible jump types. RESULTS: Table 1 shows the percentage of jumps correctly identified by the classifiers. - Sensor location 4 categories 6 jump types Waist 92.8 % 93.3 % Ankle 99.2 % 96.7 % Waist+Ankle 99.3 % 98.2 % CONCLUSION: IMU data can be leveraged in automated classification algorithms to help document jump-specific training load. Consideration should be given to the location of the sensors; while the waist+ankle was most accurate, a single sensor system may be more advantageous from the user perspective.