A Recurrent Neural Network for Hand Gesture Recognition based on Accelerometer Data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Hand gesture recognition systems often require biosignal data, necessitating affordable, reliable, and mobile solutions.
• Muscle contractions in the forearm cause skin motion, detectable by accelerometers.
• Accelerometers offer a cost-effective and mobile sensing solution for biosignal acquisition.

Purpose of the Study:

• To evaluate the feasibility of a robust hand gesture recognition system solely based on accelerometer signals.
• To propose and assess a novel neural network architecture for accelerometer-based gesture classification.

Main Methods:

• Development of a neural network architecture utilizing two distinct recurrent neural network (RNN) cell types.
• Experiments conducted on three distinct databases to validate the proposed approach.
• Classification of hand gestures using short time windows of 5 ms.

Main Results:

• The proposed RNN-based system significantly outperforms existing state-of-the-art multi-modal hand gesture recognition methods.
• The system demonstrates robust performance across different subjects and is particularly effective for amputees.
• The network achieves high accuracy with very short (5 ms) data windows.

Conclusions:

• Accelerometer data combined with RNNs provides a robust and effective method for hand gesture classification.
• This approach enables rapid and potentially real-time hand gesture detection.
• The system's affordability, reliability, and mobility make it suitable for various applications.