Author(s): MAYR, S., KRANZINGER, S.1, KRANZINGER, C.1, STOEGGL, T.2,3, Institution: SALZBURG RESEARCH FORSCHUNGSGESELLSCHAFT MBH, Country: AUSTRIA, Abstract-ID: 563

INTRODUCTION:
Big Air training relies heavily on video-based methods. While video analysis provides rich qualitative information, it requires multiple camera setups, expert interpretation, and time for processing and analysis, limiting its use during independent training sessions. Wearable inertial sensors offer a complementary approach by enabling automated, objective, and easy-to-use analysis without external infrastructure. The aim of this study is to evaluate whether jump types in elite Big Air can be classified using boot-mounted IMU data.
METHODS:
Five professional male freestyle skiers performed Big Air jumps during training sessions. IMU data were collected bilaterally using the connected boot system [1], complemented by synchronized video recordings from four cameras. Differences in jump difficulty and athlete ability resulted in a skewed class distribution within the dataset. Jump types performed fewer than four times were excluded, resulting in a dataset comprising seven jump classes (straight, 360, cork 3, cork 5, switch cork 5, cork 7, lincoln loop). From each jump, statistical time-domain features were extracted, including mean, sd, min., max., RMS, 25th, 50th and 75th percentiles, and IQR. Models emphasizing interpretability and robustness to imbalance were evaluated, including Balanced Random Forest, XGBoost, and Extremely Randomized Trees. Hyperparameters were optimized using random search with 100 iterations and group 5-fold cross-validation. Model evaluation was performed using leave-one-participant-out cross-validation. Performance metrics included balanced accuracy (BA), weighted F1-score (wF1), and Cohen's kappa (κ), reported as mean and sd across folds. Feature importance was assessed using impurity-based and permutation-importance methods.
RESULTS:
Models achieved meaningful classification performance despite strong class imbalance. The Balanced Random Forest (BA=0.72, wF1=0.69, κ=0.6) showed the most stable results, while XGBoost (BA=0.66, wF1=0.66, κ=0.6) and Extremely Randomized Trees (BA=0.75, wF1=0.69, κ=0.66) achieved comparable peak performance. Misclassifications occurred predominantly between biomechanically similar jump types, such as 360 and cork 3. Gyroscope-derived features, particularly those related to rotation around the vertical and mediolateral axes, and variability-based descriptors dominated feature importance across all models, highlighting the relevance of rotational dynamics.
CONCLUSION:
IMU data provide a promising way for classification of Big Air ski jump types in elite athletes, even under imbalanced conditions. These findings support the use of wearable sensors as a practical tool for objective jump analysis in freestyle skiing training and performance monitoring, with future potential to extend towards scoring-related quality metrics, error detection, and more targeted feedback for motor learning and performance evaluation.

[1] Snyder et al., Sensors, 2021