Pseudo-Siamese Neural Network for Bed-Exit Detection With Low-Resolution Sensors

Elderly individuals with lower limb weakness are at a high risk of falling when they are attempting to exit their beds. Thermal imaging provides privacy protection and operates independently of ambient lighting, making it suitable for detecting bed-exit events among elderly individuals in low-light conditions to alert caregivers to prevent falls. This work proposes a bed-exit alarm system that employs a low-resolution 12×16 thermopile array sensor to capture thermal images from the back of individuals in bed. A pair of thermal images in which the number of pixel changes between them exceeds a predefined threshold indicates a significant motion by the individual. These thermal images are then fed into a proposed pseudo-Siamese convolutional neural network to determine whether the motion corresponds to a bed-exit event. The proposed method achieves a detection rate of 100.0% and a false alarm rate of 1.9%, outperforming previous systems that use high-resolution thermal imaging sensors. Additionally, this system can issue an alarm when bed-exit events reach an average of 42.5% completion, which provides caregivers with lead time to offer timely assistance. The proposed system was used to monitor two nursing home residents for 60 days and 41 days, and captured 222 and 98 actual bed-exit events, respectively. These results demonstrate the feasibility of implementing bed-exit alarm systems using low-resolution sensors for elderly care.