Respiration Differentiating Using Wavelet Transform and Autoencoder

Abstract

This study proposes a novel respiration differentiation method using a wavelet transform and an autoencoder. Depth changes in a person’s chest were acquired from the depth images generated by the Kinect sensor and recorded as waveform data. The extracted waveform data were frequency-analyzed using a wavelet transform for the autoencoder learning. The autoencoder was trained exclusively on normal respiratory data. Respiratory data were identified based on the differences between the autoencoder’s inputs and outputs. To automatically differentiate respiration from other movements, a threshold was set using Hotelling’s theory. The proposed respiration differentiation method was experimentally evaluated to verify its high recognition rates, even for untrained individuals, using cross-validation. Respiration differentiation was also applied to the entire image to confirm that the model could accurately recognize chest movements resulting from respiration.