Development of an Automated Motion Evaluation System from Wearable Sensor Devices for Ski Jumping

We developed a system for the automatic evaluation of ski jumps on the base of machine learning algorithms. In several capture sessions during summer jump season, motions of four junior ski jumpers were captured by inertial sensors over the complete jump, from the start of the in-run to the end of the outrun phase. Additionally, style points were collected from an experienced judge to serve as ground truth and control data for the machine learning algorithms. All jumps were randomly separated into a training and a test database. Next, we determined kinematic factors such as body segment orientation and body joint position and segmented every jump into its main motion phases on the base of the raw sensor data. Specific motion characteristics that influence the performance and length of a jump were then used to create machine style knowledge from the training database in compliance to the official scoring guidelines. In a last step, we computed the similarity of every jump within the test database to the trained knowledge on a ski jump's style parameters and evaluated the presence or absence of jump errors. Results showed that the computed error annotations conform largely to the human-based judging scores. Adding automatic score measures to the algorithm in future, this method might be an important step towards better measurability and objectivity of performance-oriented sports in form of a mobile competition evaluation system.