A Study of Human Fitness Pose Classification Using Artificial Neural Networks
Sijie Shang, Rong Jin, Kevin Desai · 2023
Motivated by the increasing demand for detecting and classifying human poses in the realm of personalized fitness training by AI technologies, which provide feedback on the form to help users exercise more accurately, and the proven effectiveness of some deep learning models in achieving that, this study aims to investigate three different ensemble approaches for artificial neural network models for detecting and classifying human poses. The pre-trained MoveNet model was employed to extract the positions of 17 body keypoints, which were used as input data for the subsequent three classification models - a Feedforward Neural Network (FNN), LSTM, and GRU. The LSTM and GRU models have the ability to process time series data as input, resulting in improved accuracy compared to the FNN model. Specifically, the LSTM model achieved an accuracy of nearly 95%, while the GRU model outperformed with an accuracy exceeding 95% and the potential to reach 97.27 %.