Non-Contact Respiration Rate Measurement From Thermal Images Using Multi-Resolution Window and Phase-Sensitive Processing

This paper presents a method for measuring human respiration using a thermal camera. Respiration, being a commonly monitored biomedical signal, has traditionally been measured using contact-based methods, which can cause discomfort and skin damage to patients. With the need for non-contact respiration measurement to prevent infection in the post-COVID-19 era, previous studies encountered challenges in applying non-contact methods in clinical settings due to the distortion of respiratory signals caused by high-dimensional filters. To address these limitations, this paper proposes the use of a thermal camera for non-contact data acquisition and accurate respiratory rate (RR) prediction, even in the presence of noisy respiratory signals. The proposed method leverages a multi-resolution window (MW) and phase-sensitive (PS) processing. Specifically, thermal images are captured using an infrared (IR) thermal camera, and the respiratory signal is extracted from the series of thermal images. The MW method, employing three windows of different sizes in the time domain, is then applied to the extracted respiratory signal. The resulting MW output is transformed into the frequency domain using the Fourier synchro-squeezed transform (FSST). The PS processing involves the multiplication of the converted signal, which is in polar form after a phasor operation. The processed data is utilized to train two bidirectional Long Short-Term Memory (bi-LSTM) networks, and the RR is calculated based on the trained model. To validate the proposed model, a total of 37 surgical patients were involved, with 20 patients used for training and 17 patients for testing. Six deep learning models were designed and their performances were compared. The results indicate that the proposed model outperformed the others, achieving a test classification accuracy of 98.06% and a root mean square error (RMSE) of 0.381. Despite being non-invasive and non-contact, our method demonstrates high accuracy in predicting RR, attributed to the utilization of MW and PS processing. Therefore, it holds potential for application in clinical environments, such as monitoring patients with pulmonary diseases or in intensive care units (ICUs).