Microcontroller Implementation of LSTM Neural Networks for Dynamic Hand Gesture Recognition

Area of Science:

• Wearable technology
• Machine learning for human motion analysis
• Embedded systems

Background:

• Accelerometers in portable devices offer unobtrusive human motion data acquisition.
• Human motion recognition is crucial for human-computer interaction and health monitoring.

Purpose of the Study:

• To evaluate the performance of a Long Short-Term Memory (LSTM) neural network for hand-gesture classification on a microcontroller.
• To analyze the trade-offs between accuracy and resource utilization for embedded human motion recognition systems.

Main Methods:

• Utilized a publicly available dataset of 20 hand gestures from 10 subjects using wrist-worn accelerometers.
• Implemented an LSTM neural network model for gesture classification.
• Deployed and evaluated the model on an STM32L4-series microcontroller.

Main Results:

• Achieved nearly 90.25% accuracy for hand-gesture classification.
• Inference time was 418 ms for 4-second sequences.
• Average CPU usage was approximately 10% for the recognition task.

Conclusions:

• LSTM neural networks can effectively perform hand-gesture recognition on resource-constrained microcontrollers.
• The system demonstrates a viable balance between high accuracy and efficient resource utilization for embedded applications.