Deep Learning Model for Sensor based Swimming Style Recognition

This paper aim is to present deep learning based approach for swimming style recognition performed on publicly available data collected with a smartwatch. The proposed method is a Bi-LSTM (Bidirectional Long-Short Term Memory) network, which was constructed using MATLAB neural network toolbox. Data for the system was prepared by segmenting it into fixed-size windows and extracting pure signal features such as mean, standard deviation, median absolute deviation (MAD), signal magnitude area (SMA), interquartile range (IQR) as well as features from normalized signal spectrum such as entropy, energy, kurtosis, skewness and index of spectrum maximum (ISM) from each window. The Bi-LSTM method was able to perform with average F1 score of 91.39%.