Periodicity-Driven Adaptive Localization of Respiratory Regions for Non-Contact Respiration Monitoring in Sleeping Infants

Respiration is a critical vital sign for assessing the health status of infants during the first year of life. Conventional video-based respiratory monitoring methods often rely on manually defined regions of interest (ROIs) targeting specific anatomical regions, such as the chest or abdomen, or employ pretrained object detection models to localize body regions before extracting respiratory signals. However, in unconstrained sleep environments, infants may adopt diverse sleep postures, making it difficult to assume that respiration-induced movements will consistently appear in the same anatomical region or image location. To address this limitation, we propose an anchor-based respiratory-region localization framework that identifies spatial regions exhibiting prominent respiratory periodicity across the entire video frame without relying on predefined anatomical ROIs or separate body-region localization models. For each grid cell, time-series signals are constructed from inter-frame color/intensity variations and optical-flow-derived motion changes. The magnitude spectra of these signals are then multiplied element-wise to generate a respiratory periodicity map representing the spatial distribution of respiration-related periodicity across the frame. The grid cell with the highest Periodicity Score is selected as the Anchor Grid Cell and used as the primary respiratory region for respiratory-rate estimation. The proposed method was evaluated using 6-fold subject-wise cross-validation on 374 video clips from 15 infants. The single-Anchor Grid Cell approach achieved a mean absolute error (MAE) of 1.9953 breaths per minute (BPM), a root mean square error (RMSE) of 3.9681 BPM, and a Pearson correlation coefficient (PCC) of 0.7287. Moreover, MAEs below 2.5 BPM were obtained for the supine, side-lying, and prone postures, indicating consistent estimation performance across the evaluated sleep postures. These findings suggest that respiration-related information in infants is not necessarily confined to predefined anatomical regions and may instead be expressed at different spatial locations depending on sleep posture.