PULSEGUARD: Reliable Contactless Physiological Monitoring from Facial Video

This is the authors' abstract. We don't add key points for this paper.

Remote photoplethysmography (rPPG) enables contactless heart-rate estimation from facialvideo, but individual measurements can become unreliable under motion, illumination variation,and signal contamination [1], [2], [3]. Existing methods have explored chrominance-based,projection-based, and blind-source-separation approaches, while sparse spectral peak trackinghas been investigated for robust pulse-rate estimation [4], [5], [6], [7]. However, accurateestimation alone does not indicate whether a measurement should be trusted. This paper presentsPULSEGUARD, a reliability-oriented framework that augments rPPG heart-rate estimationwith signal-quality assessment, uncertainty quantification, failure-cause diagnosis, and selectiveprediction. PULSEGUARD combines six rPPG methods—Green, CHROM, POS, PCA, ICA,and LGI—through adaptive quality-weighted spectral fusion [1], [4], [5], [6], [8], [9]. Aseven-component Physiological Signal Quality Index (PSQI) characterizes spectral quality,cross-method agreement, motion, illumination, region-of-interest stability, temporal consistency,and harmonic coherence. The framework additionally estimates heart-rate uncertainty, diagnosesnine failure conditions, and assigns each measurement a REPORT, CAUTION, or ABSTAINoutcome. PULSEGUARD is evaluated on a controlled synthetic physiological video datasetcontaining normal and challenging conditions, including motion, illumination changes, occlusion,compression, resolution reduction, and frame-rate variation. Evaluation includes heart-rateaccuracy, Bland–Altman analysis, uncertainty calibration, failure detection, ablation analysis, andrisk–coverage evaluation [10]. Comparison of always-predict and selective-abstention strategiesshows that higher PSQI is associated with lower estimation error and that abstaining fromlow-reliability measurements reduces error on the accepted subset. The results demonstratethe potential of explicitly modeling measurement reliability for more trustworthy contactlessphysiological monitoring.