Toward Non-Contact Snore Detection Using Video-Based Laryngeal Motion Analysis

In this study, we present a proof-of-concept snore-event detection system that operates on audio-free RGB videos of a throat to analyze the laryngeal-prominence motion. The pipeline integrates Eulerian Video Magnification (EVM), which enhances subtle motions, and Gunnar Farnebäck vertical optical flow, which quantifies directed displacement, followed by an adaptive detector with trimmed Median Absolute Deviation (MAD) thresholds, high/low hysteresis, and a cooldown period to prevent double counting. A self-recorded mimicry-based dataset under controlled conditions was collected from a single-participant. The video clips were anonymized and annotated at the event level, producing 200 events (100 snores and 100 non-snores), and independently validated by a licensed physician. To assess the annotation reliability, the Cohen’s Kappa coefficient was κ = 0.87 (“almost perfect agreement”), supporting the clinical plausibility of the ground truth. With respect to detection performance, the full pipeline (EVM-optical flow) achieved an accuracy of 90.5%, precision of 91.8%, recall of 89.0%, and F1-score of 90.4%. Meanwhile, Wilson scores with 95% confidence intervals were also computed for proportion metrics to reflect the uncertainty of the sample size. Moreover, an ablation study confirmed that the combined method outperformed EVM-only and optical flow-only baselines, underscoring a magnify-then-gate design that is sensitive to low-amplitude throat motion and robust against non-snore artifacts. Although limited by its single-participant design, this approach demonstrates the potential for non-contact, privacy-preserving snore detection and motivates multi-subject and clinical evaluations with future integration into video-based vital-sign monitoring for broader sleep health screening.