Abstract
(1) Background: Detecting long-lie incidents-where individuals remain immobile after a fall-is essential for timely intervention and preventing severe health consequences. However, most existing systems focus only on fall detection, neglect post-fall monitoring, and raise privacy concerns, especially in real-time, non-invasive applications; (2) Methods: This study proposes a lightweight, privacy-preserving, long-lie detection system utilizing thermal imaging and a soft-voting ensemble classifier. A low-resolution thermal camera captured simulated falls and activities of daily living (ADL) performed by ten healthy participants. Human pose keypoints were extracted using MediaPipe, followed by the computation of five handcrafted postural features. The top three classifiers-automatically selected based on cross-validation performance-formed the soft-voting ensemble. Long-lie conditions were identified through post-fall immobility monitoring over a defined period, using rule-based logic on posture stability and duration; (3) Results: The ensemble model achieved high classification performance with accuracy, precision, recall, and an F1 score of 0.98. Real-time deployment on a Raspberry Pi 5 demonstrated the system is capable of accurately detecting long-lie incidents based on continuous monitoring over 15 min, with minimal posture variation; (4) Conclusion: The proposed system introduces a novel approach to long-lie detection by integrating privacy-aware sensing, interpretable posture-based features, and efficient edge computing. It demonstrates strong potential for deployment in homecare settings. Future work includes validation with older adults and integration of vital sign monitoring for comprehensive assessment.