Kinect-based shoulder strength training recognition and correction
Kunpeng Shen, Wenlu Yang · 2025
More and more people are opting to improve their physical fitness and develop the perfect body shape as a result of society's growing concern for health and form maintenance. However, improper motions during the real fitness process can easily reduce the benefits of exercise and even cause joint and muscle injuries, which might pose health problems. As a result, one of the key areas of modern scientific fitness study is how to precisely identify and rectify fitness movements using technical methods. This research proposes an efficient fitness action identification system based on the Kinect device's ability to collect human skeletal point data. First, using human kinematics theory, 17-dimensional distance and angle features are chosen to depict the shoulder training trajectory. This paper employs the Holt filter and Kalman filter, respectively, to smooth the movement trajectory and denoise the data in order to guarantee the correctness and stability of the experiment's data. After comparison, the Holt filter is more suited for processing the data in this paper. This research then trains LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) recurrent neural networks on the processed dataset and compares the results. According to the experimental data, the GRU-based model outperforms the LSTM-based model in terms of recognition accuracy, reaching 97.72%.