Highly-accurate gesture recognition based on ResNet with low-budget data gloves

Abstract

Gesture recognition has been widely used in various fields such as virtual reality (VR), human-computer interaction, and clinical medicine. Traditionally, gesture recognition requires expensive sensors to obtain high-precision data, but some applications-such as VR games-do not require high accuracy and users prefer a lower cost instead. Thus some researchers began to use Convolutional Neural Networks (CNN) to recognize gestures directly without other calibration equipment. However, using CNN encounters the degradation phenomenon of the neural network. In this paper, we propose a highly-accurate gesture recognition method for low-precision glove data based on Residual neural Networks (ResNet). Our method combines 1D convolution with 2D convolution and adopts a sliding window for sample size constraint to achieve data enhancement. Evaluation results indicate that compared to the CNN method, our method improves the classification accuracy rate from 94.4% to 99.2%.