Excited ICU Delirium Behavior Detection Method Based on Human Key Points Recognition

Currently, existing human key point detection algorithms have poor adaptability and are easily affected by light intensity in specific scenarios such as delirium early warning in the intensive care unit. Therefore, these algorithms cannot achieve 24/7 monitoring identification. To address these issues, this paper proposes a delirium early warning algorithm based on human key points estimation under infrared vision. The data input into the algorithm is infrared images, which effectively avoids the influence of light intensity on the algorithm's performance and enables 24/7 monitoring detection. The model is lightweight based on the original OpenPose key point recognition algorithm. Simultaneously, the improved OpenPose algorithm achieves a dual benefit: it streamlines parameter usage while enhancing key point recognition accuracy. This paper also proposes a soft mechanism of delirium recognition based on human skeleton sequence, which implements velocity constraints on the identified key points in the time domain to assist delirium recognition. The proposed method was verified by experiments on the self-built infrared vision key point detection dataset of human lying posture. The average recognition accuracy of all key points is 77.3%, and the average recognition accuracy of the improved OpenPose algorithm is 4.9% higher than that of the original algorithm. Based on the human skeleton sequence recognized by the algorithm, it can identify and warn of delirium behavior in severe patients.