Hand Gesture Classification based on Short-Time Fourier Transform of Inaudible Sound

As IoT devices and wearable devices are increasing in number, researches for recognizing and classifying human action or behavior as a method for controlling them have become important. This study proposes a method to classify hand gestures by one smartphone using high frequency sound which is inaudible. While generating the sound from the smartphone and conducting hand gestures, we have classified the reflected sound. The reflected data differs between the hand gestures because of the Doppler effect and we have analyzed this data in the time domain and frequency domain with short-time Fourier transform. In this paper, we have presented a convolutional neural network model that classifies 8 hand gestures, and the model showed 94.25 % classification accuracy.