Hand Gesture Recognition Using EMG-IMU Signals and Deep Q-Networks

Area of Science:

• Robotics and Human-Computer Interaction
• Machine Learning and Artificial Intelligence
• Biomedical Engineering and Signal Processing

Background:

• Hand gesture recognition (HGR) systems commonly use supervised learning with electromyography (EMG) and inertial measurement unit (IMU) signals.
• Reinforcement learning (RL) presents a promising alternative for HGR, offering advantages in classification performance and online learning capabilities.

Purpose of the Study:

• To develop and evaluate a novel HGR system utilizing RL for classifying static and dynamic hand gestures.
• To compare the performance of RL-based HGR using different sensor types (Myo armband and G-force).

Main Methods:

• An RL-based agent employing a deep Q-network (DQN) algorithm was developed for gesture classification.
• The system integrated EMG and IMU data, processed through pre-processing, feature extraction, classification, and post-processing stages.
• A feed-forward artificial neural network (ANN) represented the agent's policy, trained and tested on user-specific models.

Main Results:

• The RL-based system achieved high accuracy, reaching up to 97.50% for static gesture classification and 98.95% for dynamic gesture classification using the Myo armband sensor.
• Recognition accuracy for static and dynamic gestures reached 88.15% and 90.47%, respectively, with the Myo armband sensor.
• The study demonstrated the efficacy of RL methods, specifically DQN, in learning from online experience for HGR.

Conclusions:

• Reinforcement learning, particularly the DQN algorithm, is a viable and effective method for hand gesture recognition using EMG and IMU signals.
• The developed RL system demonstrates robust performance in classifying and recognizing both static and dynamic hand gestures.
• The findings highlight the potential of RL for creating adaptive and accurate HGR systems.