Machine learning and real-time prediction of human motion for intelligent human-machine interfaces

• Main Author: Chen, Yongming
• Other Authors: Lin Zhiping
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: https://hdl.handle.net/10356/154403
• Institution: Nanyang Technological University

With the rapid development of human-computer interaction, researchers are extending beyond physical-motion means by detecting as early as possible neurophysiological signatures of motion intentions to detect human movement and predict human movement intention. The use of electromyogram (EMG) for human movement prediction has become a major research & development field with successful examples in hand-gesture recognition, gait detection. Yet major open questions remain especially on the EMG dynamics of motion intention ever before any physical movement has started, and also on the major factors affecting the dynamic patterns. This work presents our machine learning-based approach to studying the EMG dynamics for motion intention prediction. Particularly, we designed a hierarchical motion recognition algorithm based on a Hidden Markov Model. The pre-intention muscle tone state is determined from EMG, then a respective motion-intention recognition model is selected for motion decoding. The results show that the pre-intention muscle tone states recognition accuracy of the model is 79.91%±15.94, which is higher than the method for comparison. Also, different pre-intention muscle tone states (relaxed or prepared) have an impact on the recognition accuracy of the motion intention. On average, the accuracy is 95.58% ± 3.26, in the relaxed state, versus 90.45%±8.77 in the prepared state. This is higher than the accuracy achieved using traditional HMM training methods. We also find that for early motion prediction, the accuracy of motion recognition is correlated with the proportion of hidden states in a specific stage of the HMM model. This work suggests that a proper design of HMM-based motion intention recognition can improve human-machine interfaces and provide new insights into the process of human movement intention.