Deep-Learning Technique for Bed-exit Action Prediction of Elderly Using Extremely Low-resolution Thermopile Sensor Array

Elderly Care is becoming an increasingly important issue in aging and super-aged societies. In particular, Taiwan’s Long-Term Care 3.0 policy highlights “smart assistive technologies” as one of its key areas for promotion. A critical focus in elderly care is fall prevention, as falls in older adults often result in serious or even fatal injuries. Given that bedrooms and bathrooms are the two primary hotspots for falls, this study focuses on the former. The core objective of this research is to prevent falls. However, predicting when a fall will occur is extremely difficult—even in healthy individuals (e.g., someone might trip unexpectedly). Therefore, we considered the typical routines and potential fall causes among elderly individuals and identified a high-risk period during the transition out of bed. During this period, conditions such as orthostatic hypotension or musculoskeletal weakness may lead to momentary dizziness and result in falls. To address this issue, our approach transforms the goal of “fall prevention” into a strategy of “early bed-exit prediction.” By detecting—or even predicting—the elderly person’s intention to get out of bed, our system can alert caregivers, who can then assist verbally (e.g., calling out) or physically, thereby preventing potential falls. To provide caregivers with sufficient response time, this study emphasizes bed-exit prediction rather than simple detection, allowing a buffer of 5–8 seconds for verbal or physical intervention. Currently, smart assistive devices in many care centers rely on air cushions or pressure mats (placed under mattresses) that detect movement only after the person has already gotten out of bed. This is often too late to prevent a fall. Our technology, by contrast, provides an alert during the critical 5–8 second window, significantly improving the chances of fall prevention. Because cameras often raise privacy concerns, our study uses a thermopile-type far-infrared sensor with very low resolution (32×32 pixels). This sensor is inexpensive and inherently privacy-preserving, as the images it produces are unrecognizable to the human eyes. We apply cutting-edge AI and deep learning technologies to predict bed-exit intent on average 5.78 seconds before the actual action. Our database tests show 99.37% accuracy and 99.09% precision. The system supports real-time, online, continuous operation, not just offline database testing, placing it at the forefront of international research. The thermopile sensor used in this study does not present the privacy risks associated with RGB color cameras and is affordably priced between NT$1,500–10,000. However, due to its low resolution (32×32 pixels), traditional image processing and computer vision algorithms perform poorly. To overcome this, we adopted the latest AI and deep learning techniques for practical effectiveness in action prediction. We installed 1–3 far-infrared sensors above the bed (non-intrusive to daily activities) to continuously capture thermal imagery. The data is integrated to anticipate the senior’s intent to get out of bed, allowing caregivers to assist before a fall might occur due to orthostatic hypotension or insufficient leg support. Our system utilizes a combined CNN/LSTM/GRU neural network, along with a custom early-prediction loss function, to accurately predict the senior’s intent to get up—without triggering false alarms caused by minor movements like turning over or lifting a leg during sleep. The results of this research have been showcased at our university’s CIRAS center (Center of Innovative Research for Aging Society) on several occasions, leaving a strong impression on visitors. The findings were also published in 2023 in the IEEE Transactions on Circuits and Systems for Video Technology, an SCI Q1 journal (Ranking: 91.1%) 圖 1 本研究熱電堆紅外線 (TPA) 影像實拍圖 圖 2. 本研究三個紅外線影像感測器 (TPA) 實際配置圖 (床上天花板處) 圖 3. TPA 紅外線影像動作預測系統介面 (左上為三個 TPA 影像，右面為估測之離床機率曲線之時間變化及使用之紅色門檻線(threshold), 當超越門檻線，及發出告警。   Research team：Intelligent Signal Processing, ISP: 520 多媒體實驗室 賴文能老師團隊 (Lie group), Lie group team, R520 laboratory of Intelligent Signal Processing (ISP) group, Department of Electrical Engineering, National Chung Cheng University, Taiwan