MEMS Devices-Based Hand Gesture Recognition via Wearable Computing.

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Gesture recognition is vital for VR, medical diagnosis, and robotics.
• Current methods include inertial-sensor-based and camera-vision-based approaches.
• Optical methods face limitations like reflection and occlusion.

Purpose of the Study:

• Investigate static and dynamic gesture recognition using miniature inertial sensors.
• Evaluate machine learning algorithms for static gesture recognition.
• Assess Hidden Markov Models and attention-based LSTMs for dynamic gesture recognition.

Main Methods:

• Hand-gesture data collected via data glove, preprocessed with Butterworth low-pass filtering and normalization.
• Magnetometer correction using ellipsoidal fitting and auxiliary segmentation for data processing.
• Static gesture recognition employed Support Vector Machine (SVM), Backpropagation Neural Network (BP), Decision Tree (DT), and Random Forest (RF).
• Dynamic gesture recognition utilized Hidden Markov Models (HMM) and Attention-Biased Mechanisms for Bidirectional Long- and Short-Term Memory Neural Network Models (Attention-BiLSTM), compared against traditional LSTM.

Main Results:

• Random Forest achieved the highest accuracy and fastest recognition time for static gestures.
• The attention mechanism significantly improved LSTM model accuracy for dynamic gestures, reaching 98.3% with a six-axis dataset.
• Analysis revealed differences in accuracy for complex dynamic gestures based on feature datasets.

Conclusions:

• Miniature inertial sensors offer a viable alternative for gesture recognition, overcoming optical limitations.
• Random Forest is highly effective for static gesture recognition tasks.
• Attention mechanisms substantially enhance the performance of deep learning models in dynamic gesture recognition.