Swimming Style Recognition with Convolutional Neural Network with Single IMU

Abstract

The study focuses on swimming style recognition using a Convolutional Neural Network (CNN) and a single inertial measurement unit (IMU) located on the wrist, mimicking the setup of a typical smartwatch. Traditional methods often employ multiple sensors or sensors placed in impractical locations, limiting everyday usability for swimmers. This research utilizes data collected from 40 swimmers, integrating 53,732 input windows from sessions in a 50-meter pool, where swimmers wore a smartwatch on their wrist. The CNN model was modified to enhance pattern recognition capabilities across swimming styles: Butterfly, Backstroke, Breaststroke, Freestyle, and Transitions. Enhanced data processing techniques, including normalization and augmentation (time-scaling, noise addition, reversing, and rotation), were applied to simulate real-world variances. The model was evaluated using a Leave-One-Subject-Out (LOSO) cross-validation method, demonstrating high precision and minimal misclassification, signifying a robust model capable of accurately detecting and classifying swimming strokes with potential for real-time application in sports and health monitoring. This approach offers a significant improvement over previous systems by reducing the sensor complexity and focusing on individual movement patterns, potentially increasing the accessibility and precision of swimming activity recognition.