[Reliability-aware periodicity graph neural network for remote heart rate estimation]

Remote photoplethysmography is susceptible to phase perturbations and spatiotemporal noise under complex motion conditions. Conventional periodic modeling methods typically assign equal weights to graph-structured temporal segments, thereby overlooking differences in node reliability and temporal variations in signal quality. To address this limitation, this paper proposes a reliability-aware periodic graph neural network for robust rPPG estimation. First, a periodic consistency constraint module is introduced to evaluate the stability of temporal segments based on cross-period phase-alignment relationships. Second, a spatio-temporal reliability modulation module is designed to jointly characterize feature quality from two complementary perspectives: spatial-region confidence and temporal-difference stability. This enables explicit modeling of node reliability and collaborative optimization of spatiotemporal features. Experimental results demonstrate that, compared with the baseline model, the proposed method reduces the mean absolute error from 0.55 to 0.27 on a public pulse estimation dataset and from 1.61 to 0.07 on the University of Burgundy Franche-Comté dataset, validating the effectiveness and stability of the proposed reliability-aware mechanisms under complex motion conditions. Overall, the proposed method provides an effective reliability-enhancement strategy for periodic modeling in rPPG under challenging interference conditions.