|
|
|
| I. Research objective According to the "2021 Taiwan Patient Safety Report," 20.9% of fall incidents in hospitals occur during the process of getting in and out of bed, with most victims being elderly individuals. To prevent falls, caregivers are asked to assist elderly patients when they try to get out of bed. However, due to Taiwan’s severe nurse-to-patient ratio (1:13), it is difficult for a single caregiver to monitor multiple elderly patients at once, posing a significant challenge to fall prevention. Existing commercial products that detect bed-exit motions, such as pressure-sensing mats and cameras, have limitations. Pressure-sensing mats can easily produce false alarms due to the patient's weight and on-bed motions, while cameras capture RGB images that raise privacy concerns for the elderly. Therefore, a non-contact, privacy-preserving bed-exit detection method is necessary for fall prevention and the safety of elderly individuals. We have developed a bed-exit detection system using thermal imaging combined with machine learning models to recognize the bed-exit motions of elderly individuals. The system sends out an alert at the early stage of bed-leaving actions, allowing healthcare personnel to assist the elderly promptly. Additionally, frequent false bed-exit alarms may lead to alarm fatigue, delaying caregiver responses to actual emergencies. Therefore, the proposed system integrates augmented reality (AR) glasses to enable caregivers to remotely view live images of the elderly when an alarm is triggered. This mechanism allows caregivers to immediately determine whether the senior needs support and prevents false alarms from interfering with their work. |
|
Figure 1. Current issues in bed-exit detection and proposed solutions by our team |
|
| II. Research content Thermal Image-based bed-exit detection system: The proposed system (Fig. 2) uses a low-resolution thermal imaging sensor with a resolution of 12x16, mounted at the head of the bed to capture thermal image sequences during the bed-exit process of elderly individuals. This installation position avoids interference from blankets and passing people. The sensor can effectively detect a human body within two meters. Compared to cameras and depth sensors, the low-resolution thermal imaging sensor is cost-effective, unaffected by background light noise, and ensures the privacy of the elderly in their daily lives, increasing the acceptance of the proposed system by both the elderly and their family members. |
|||
Figure 2. Thermal Image-based bed-exit detection system |
|||
|
Algorithm:
The proposed system is designed to detect elderly individuals getting out of bed from either side of a hospital bed. In thermal images, the boundary pixels between the background and the foreground parts can be considered as the contour of the human body. Human motions can cause temporal changes in this contour. By analyzing these changes, bed-exit motions can be identified. In this study, the proposed system extracts the contour changes from N consecutive thermal images. It's noteworthy that the temperature of a blanket covering the body is similar to that of the boundary pixels, and moving the blanket can also cause changes in the contour, increasing the complexity of bed-exit detection. This complexity makes simple threshold-based methods ineffective for classifying human motions. Thus, in this study, a neural network based on 1D convolutional neural network (1D CNN) layers is used to extract features of the contour changes to identify bed-exit motions. Fig. 3 illustrates the model architecture, where contours from N frames of thermal images are input to output a bed-exit score (0-1, with scores above 0.7 considered as bed-exit). Table 1 shows the effect of different N values on the detection, where the latency is defined as the time between the start of a bed-exit motion and the detection by the model. When N is small, the model predicts bed-leaving based on the contour changes within a short duration, resulting in short latency; as N increases, the model uses long data sequences for detection. Experimental results indicate that with N=3, the model achieves a latency of 3.02 seconds and a sensitivity of 98.8% for detecting bed-leaving actions. |
|||
Figure 3. Proposed 1D CNN-based bed-exit detection model. |
|||
Table 1. Effect of N on bed-exit detection |
|||
|
AR glasses and real-time streaming:
In addition to the thermal imaging sensor, the system includes a camera connected to AR glasses worn by the caregiver to transmit real-time images of the elderly (Fig. 4). To protect the privacy of the elderly, the camera is covered by a shutter controlled by a Raspberry Pi when there is no bed-exit alarm. Only the thermal imaging sensor is active for bed-exit detection. In the absence of a bed-exit alarm, the camera remains off, and the AR glasses display a transparent screen to avoid obstructing the caregiver's view and affecting their daily tasks. When the system detects a bed-exit motion through thermal images, the Raspberry Pi is notified via a Zigbee module to remove the shutter and activate the camera, then transmits the elderly's real-time image to the AR glasses. With the real-time images, caregivers can immediately assess the elderly’s condition, avoiding unnecessary responses to false alarms. When the bed-exit alarm ends, the camera is covered again, protecting the privacy of the elderly in their daily lives. |
|||
Figure 4. Bed-exit detection system integrated with AR glasses and real-time streaming. |
|||
