Remote Recognition of In-Bed Postures Using a Thermopile Array Sensor With Machine Learning

Abstract

Recently, there has been growing attention to sleep quality monitoring. In-bed posture is not only one of the critical factors that evaluates sleep qualities but can also help improve the well-being of bedridden patients. Our previous work has reported the successful recognition of 5 in-bed postures using a wall-mounted thermopile array sensor with the hand-crafted feature extraction-based machine learning. In this article, both ceiling-mounted and wall-mounted thermopile array sensing approaches are devised and compared to identify 9 typical in-bed postures using advanced machine learning models. In particular, a preprocessing method and a synergistic feature extraction approach using integrated histogram of oriented gradient and principal component analysis are proposed. An accuracy of over 99.8% is reached when using the 5-fold cross-validation to classify 9 postures. Besides, the leave-one-subject-out cross-validation (cross-user-validation) is performed to evaluate and compare the robustness of the classifiers using both sensor-mounting approaches. As the thermopile array sensor has the unique property of non-contact, non-privacy invasion, and passive sensing, the proposed system provides a contactless, unobtrusive, low cost and convenient solution for long-term sleep and in-bed patient monitoring.