Thermal Image-Based Bed Monitoring System Using VLP Model for Elderly Fall Prevention

• Main Authors: Hsieh, Ping-Hung, Lee, Po-Ting, Yang, Jia-Han, Sung, Pi-Shan, Lin, Chih-Lung
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Bed Monitoring System; Image recognition; Monitoring; Older adults; Posture recognition; Recording; Safety; Sensor systems; Thermal sensors; Transformers; Vision-Language Pretraining (VLP); Visualization
• ISSN: 2168-9229
• DOI: 10.1109/SENSORS60989.2024.10785123

Falls occur when elderly individuals get out of bed alone and can cause serious injuries to them. Bed monitoring systems monitor the postures of the elderly near the bed to prevent them from falls. Integrating bed monitoring systems with electronic health records (EHRs) aids caregivers in assessing the elderly's risk of falls. This work proposes a bed monitoring system that uses a Vision-Language Pre-training (VLP) model to generate captions of a person's posture in bed for EHRs. The proposed system uses thermal images to identify a person's posture without compromising privacy. These thermal images are then converted into visual representations with a fine-tuned vision transformer and fed into the VLP model to obtain captions. The proposed system achieves 100% accuracy in generating captions for 770 thermal images across ten postures from ten participants, demonstrating its capability to recognize human postures and generate recordings for EHRs, thereby enhancing safety for the elderly.