SIAR: Signing in the Air Based on Patching-Gated Recurrent Unit Model for Fine-Grained Trajectory Feature Learning
Shih‐Hsiung Lee, Hsuan-Chih Ku · IEEE Sensors Journal · 2025
In the current era, where information security and personal privacy are paramount, biometric recognition technology has raised significant privacy concerns. In particular, the COVID-19 pandemic has accelerated the demand for contactless biometric recognition technology, among which facial recognition technology is widely used owing to its convenience, but also faces privacy protection controversies. To address this challenge, this study innovatively proposes a noncontact identity verification method based on hand-air–writing trajectories. This method avoids reliance on personal facial data and reduces the risk of privacy infringement. It also significantly reduces the chance of virus transmission as it requires no contact with any surface, providing a secure and privacy-protective new option for biometric recognition technology. This study adopted the gated recurrent unit (GRU) model to process and analyze the hand trajectory data. The GRU model was selected for its excellent efficiency in handling time-series data and for effectively extracting distinctive features from user hand gestures. Applying this model enabled the system to accurately identify different symbols written in the air by users, thereby completing the identity verification process. To achieve wider application and reduce system operating costs, this study deployed a model on edge-computing devices. At a precision level of fp32, the system achieved an accuracy of 97.33%, demonstrating considerable practical potential.