inis 24(3):

Research Article

Resource-Efficient Deep Learning: Fast Hand Gestures on Microcontrollers

Download34 downloads
  • @ARTICLE{10.4108/eetinis.v11i3.5616,
        author={Tuan Kiet Tran Mach and Khai Nguyen Van and Minhhuy Le},
        title={Resource-Efficient Deep Learning: Fast Hand Gestures on Microcontrollers},
        journal={EAI Endorsed Transactions on Industrial Networks and Intelligent Systems},
        volume={11},
        number={3},
        publisher={EAI},
        journal_a={INIS},
        year={2024},
        month={7},
        keywords={Human-Computer Interaction, Microcontroller, TinyML, Intelligent system, Hand gesture recognition, Embedded, Communications},
        doi={10.4108/eetinis.v11i3.5616}
    }
    
  • Tuan Kiet Tran Mach
    Khai Nguyen Van
    Minhhuy Le
    Year: 2024
    Resource-Efficient Deep Learning: Fast Hand Gestures on Microcontrollers
    INIS
    EAI
    DOI: 10.4108/eetinis.v11i3.5616
Tuan Kiet Tran Mach1, Khai Nguyen Van1, Minhhuy Le1,*
  • 1: Phenikaa University
*Contact email: leminhhuy8886@gmail.com

Abstract

Hand gesture recognition using a camera provides an intuitive and promising means of human-computer interaction and allows operators to execute commands and control machines with simple gestures. Research in hand gesture recognition-based control systems has garnered significant attention, yet the deploying of microcontrollers into this domain remains relatively insignificant. In this study, we propose a novel approach utilizing micro-hand gesture recognition built on micro-bottleneck Residual and micro-bottleneck Conv blocks. Our proposed model, comprises only 42K parameters, is optimized for size to facilitate seamless operation on resource-constrained hardware. Benchmarking conducted on STM32 microcontrollers showcases remarkable efficiency, with the model achieving an average prediction time of just 269ms, marking a 7× faster over the state-of-art model. Notably, despite its compact size and enhanced speed, our model maintains competitive performance result, achieving an accuracy of 99.6% on the ASL dataset and 92% on OUHANDS dataset. These findings underscore the potential for deploying advanced control methods on compact, cost-effective devices, presenting promising avenues for future research and industrial applications.