ThermRNet: Deep Spatiotemporal Learning for Non-Contact Respiratory Rate Estimation via Infrared Thermography

Respiratory rate is a vital physiological indicator, yet traditional monitoring methods often require physical contact, limiting their usability in unobtrusive or remote settings. Infrared thermography enables non-contact respiratory monitoring but faces challenges in robustness and accuracy under complex conditions. To address these issues, we propose ThermRNet, a novel deep spatiotemporal learning model designed for reliable and precise respiratory monitoring using thermal video data. ThermRNet fuses 3D convolutional neural networks (3D-CNN) for localized spatial-temporal feature extraction with a Transformer-based spatiotemporal attention mechanism that captures long-range dependencies across frames. This architecture enables the model to both recognize instantaneous breathing states on a per-frame basis and accurately estimate the continuous respiratory rate over time. On a thermal video dataset collected from 15 participants, our model achieves 95.1% classification accuracy, a mean absolute error of 0.2166 bpm, and a Pearson correlation coefficient of 0.9958. These results demonstrate ThermRNet's strong potential for reliable, contactless respiratory monitoring in real-world applications such as telemedicine and intelligent health surveillance applications.