Recognition of hand action using body-conducted sounds

Several methods of recognizing hand actions by examining the vibrations conducted through the body from muscular activity (we call them “body-conducted sounds” in this paper) were proposed in previous works. However, they did not consider the transfer characteristic of body-conducted sounds. In this paper, we propose a method for hand action recognition that extracts the main frequency elements of body-conducted sounds using the Mel-Frequency Cepstrum Coefficient (MFCC) and divides the hidden states of Hidden Markov Models (HMMs) based on the MFCC. The results of experiments show that our method makes it possible to correctly recognize 95% of hand actions on average.