First-hand information for everyone
Researchers have shown that an end‑to‑end deep‑learning system can predict respiratory frequency directly from raw thermal video, eliminating the need for laborious hand‑crafted regions of interest. The approach uses a 3‑dimensional convolutional neural network (3‑D‑CNN) that ingests 30‑second video clips sampled at 5 Hz and outputs a breathing‑rate estimate in breaths per minute.
Because the available data set is small, the authors first trained the network on a large synthetic data set created by superimposing pulsating circles onto static facial backgrounds. This synthetic transfer learning step provided a strong starting point for later fine‑tuning on real recordings. They further employed RandAugment to randomly alter the video frames—adding noise, blur, distortion, and rotation—to increase variability and reduce overfitting.
The combined training strategy achieved a validation coefficient of determination (R²) of about 0.61 on both raw and pre‑processed thermal videos. In the raw‑video setting the lowest mean‑squared error (MSE) was 0.312 and the best R² was 0.604, while on pre‑processed videos the corresponding numbers were 0.304 MSE and 0.614 R². When the model was evaluated across all subjects in a 6‑fold cross‑validation, it maintained a mean R² of 0.663 (± 0.02 SD) and a mean MSE of 0.297, confirming consistent performance.
The work demonstrates that a small, data‑augmented, deep‑learning model can infer breathing patterns from thermal imagery, a promising step toward contactless respiratory monitoring. However, the study’s protocol limited respiration rates to 9–30 cycles per minute and involved 30 stationary subjects. No head motion or physical activity was included, so real‑world scenarios such as driver monitoring—where people move and breathe faster—remain untested.
Future work may broaden the data set, incorporate motion robustness, and evaluate the model on more varied respiratory rates to strengthen generalisation for practical applications.