A Rapid Adaptation Approach for Dynamic Air-Writing Recognition Using Wearable Wristbands with Self-Supervised...

Area of Science:

• Wearable technology
• Deep learning
• Human-computer interaction

Background:

• Existing hand gesture recognition systems often struggle with dynamic movements and require large labeled datasets.
• There is a need for adaptable and efficient gesture recognition systems for daily activities.

Purpose of the Study:

• To develop a wearable wristband system for accurate dynamic hand gesture recognition.
• To enable rapid adaptation to various tasks using self-supervised learning and few-shot fine-tuning.

Main Methods:

• A four-channel sensing array with ionic hydrogel and flexible electrodes for high-sensitivity capacitance output.
• A deep learning algorithm employing self-supervised contrastive learning to extract features from unlabeled wrist movement data.
• Wireless transmission via a Wi-Fi module for real-time signal processing.

Main Results:

• The system achieved 94.9% accuracy in recognizing eight-direction commands and air-writing all numbers and letters.
• Demonstrated rapid adaptation to new tasks with only few-shot labeled data required for fine-tuning.
• Enabled seamless task transitions without structural modifications or extensive retraining.

Conclusions:

• The proposed wearable wristband system offers a novel approach to dynamic hand gesture recognition.
• The self-supervised learning method significantly reduces the need for labeled data, enhancing adaptability.
• The system provides a natural and intuitive interface for enhanced human-machine interaction across diverse applications.