Detecting Rowing Strokes Faults With a Single On-Body Inertial Sensor

Since their introduction, inertial navigation systems have brought innovation in many domains like, e.g., robotics, avionics, and automotive, where the combination of accelerometer, gyroscope, and sometimes magnetometer sensors has resulted helpful. More recently, the introduction of Inertial Measurement Units (IMU) with their large availability, low costs, and small dimensions, have broadened such innovations to many other fields as, e.g., health and sports. For sports performance, IMUs have opened new scenarios allowing for in-field analysis of large amount of signals data, with little burden to athletes activity. In this paper, we present a method exploiting machine learning to recognize rowing strokes characterised by premature trunk engagement, a common rowing technique error, with only one IMU positioned on the athlete lower trunk. IMU signals coming from the 9 recruited rowing athletes were first segmented into 273 strokes, translated into 22 numerical features (analytically computed from signals), and then labeled as technically correct or incorrect according to the presence of premature trunk engagement, using stereophotogrammetry data, following biomechanical and analytical considerations. Correct strokes were 134 while incorrect ones were 139. Finally, we built several classification models, exploiting automatic features selection and Leave-One- Subject-Out double cross validation for hyperparameters tuning. Our results suggest the adoption of a support vector machine (SVM) model, with accuracy of 77.3% and F1-score of 78.9%, thus showing the possibility to automatically detect rowing strokes characterised by premature trunk engagement with a single wearable IMU, with 2 features selected.