A Multimodal Recognition on Human Lying Posture

In order to assist the posture management of patients on a nursing bed, a lying posture recognition method is proposed. By using MobileNetV3 network, we fusion the information from both a lying pressure array with 8×16 sensing cells and a camera to estimate the patient’s lying posture. We set up a lying posture dataset of 15 samples with 3100 images to train the network and verify identification accuracy. We define six clusters of lying postures for the dataset, including supine, left-sided, right-sided, prone, sitting, and nobody. It was experimentally demonstrated that the recognition accuracy based on RGB image plus pressure image reached 99.29%, and the accuracy of gray image plus pressure image reached 98.99%.