GESTURE-BASED CONTROL FOR LOW-END SMART DEVICES: A HAND POSE RECOGNITION SYSTEM

Abstract

Smart devices have become ubiquitous in various aspects of modern life, including work, entertainment, shopping, and education. While touch screens have revolutionized the way we interact with mobile devices, future devices such as Google glasses and smart watches are expected to offer more interactive and intuitive interfaces. In noisy environments or for individuals with hearing impairments, voice recognition may not be an effective control solution, highlighting the importance of in-air gesture recognition for device control. Developing an interactive system for mobile devices presents numerous challenges, especially for low-end devices with limited computational capabilities. This study aims to design and develop an in-air hand pose detection system for low-end smart devices without the need for additional peripherals, relying solely on the integrated camera.

The main challenge in this context is managing a large number of hand poses, particularly on low-end devices with limited resources. While various algorithms have shown efficiency in human-computer interaction, achieving a balance between realism and robustness under different lighting conditions remains difficult, especially on low-end devices. The computational limitations of entry-level smartphones and the increased complexity of vision-based recognition tasks pose further challenges. Additionally, the introduction of Android OS Go edition for devices with limited RAM exacerbates the difficulty of performing vision-based tasks on these devices.

This work focuses on developing a system that works well on Android devices, considering their widespread usage and affordability compared to iOS devices. To address the computational limitations of low-end devices, this study employs Histogram of Oriented Gradient (HOG) features and Support Vector Machine (SVM) classification. The proposed system achieves a recognition rate of approximately 94%, outperforming existing systems based on Random Forest and back propagation neural network classifiers. The main contributions of this research include improved recognition rates compared to previous studies and the identification of cost-effective calculation methods that do not impact smartphone performance