A hybrid CNN-spectral architecture for non-contact respiratory rate estimation using multi-region optical-flow analysis

Respiratory rate (RR) is a key indicator for assessing health conditions, yet noncontact measurement remains challenging due to motion artifacts, lighting variability, and skin-tone differences. This study presents a robust framework combining a custom convolutional neural network (CNN) with spectral analysis of optical-flow signals to estimate RR accurately across diverse population. Respiration-induced motion is extracted from six anatomical regions: forehead, cheeks, upper chest, and shoulders. Adaptive Kalman filtering combined with signal-to-noise ratio (SNR)-based weighted fusion enables reliable RR estimation. To improve generalizability, extensive data augmentation was applied, simulating illumination conditions ranging from normal to low light. The experimental results indicate that the proposed method achieves a mean average error (MAE) of 0.61–0.95 breaths per minute (bpm) for different skin tones and ages, within the clinically relevant range. These findings support the effectiveness of the multi-region CNN-spectral framework as a reliable, noncontact, real-time respiratory monitoring solution with potential for continuous healthcare and telemedicine applications.