Time-Frequency Analysis and Setting Motion Activity Detection of Volleyball Motion Signals for Set-Play Classification

Human Activity Recognition with the use of wearable devices and sensors has found application in several domains including fitness, training, and sports. Various digital signal processing techniques have been applied to sensor signals for improved performance of models in recognizing and classifying sports activities. Recent studies in sports, particularly in volleyball focused on a player's spiking motion. This study focuses on the recognition of a volleyball player's setting motion. Time-frequency analysis of motion signals is performed acquiring insights on the design and application of a digital filter in capturing the relevant signals that characterizes the flick of the wrist when setting a ball. The preprocessing technique is labeled as a setting motion activity detection. Time and frequency domain features were derived from the filtered signals and fed to three deep learning models namely, Edge Impulse, custom ANN, and custom LSTM to classify set-plays into three categories – open sets, middle sets, and opposite sets. The Edge Impulse, ANN, and LSTM models yielded a 95.11%, 96.23%, and 91.67% accuracy respectively.