A Frequency Loss Driven Framework for Respiration Monitoring Using Depth-Sensing Cameras

Respiratory monitoring is crucial for detecting physiological distress, yet traditional contact-based methods are often intrusive and impractical for long-term use. Non-contact approaches, particularly camera-based techniques, offer a promising alternative, but conventional RGB-based systems raise concerns regarding data complexity and privacy. Depth-sensing cameras, by contrast, inherently safeguard privacy and simplify data handling, while remaining robust to lighting variations. However, challenges such as subtle respiratory motion, occlusions, and noise continue to reduce the effectiveness of depth-based respiration monitoring. To address these limitations, this research presents a deep learning framework tailored for breathing signal extraction and respiratory rate prediction solely from depth images. The proposed architecture combines a Convolutional Neural Network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) network to model temporal dynamics. A novel frequency domain loss function further guides the training process, encouraging accurate estimation of the dominant respiratory frequencies. The model was trained and evaluated using the pub licly available ”Breathing In-Depth” dataset, encompassing diverse breathing rates, postures, and subject-camera distances. Experimental results demonstrate that the frequency-optimized model significantly outperforms traditional time-domain training approaches, achieving superior respiratory rate accuracy while maintaining resilience against noise and occlusions. These findings highlight the potential of depth-only, frequency-driven frameworks for robust, privacy-preserving respiratory monitoring.