A Deep Learning-Based Method for Nighttime Sleep Posture Recognition

Sleep is indispensable for human health. Sleeping posture is one of the factors that affect the quality of sleep. Besides, correct sleeping postures could prevent and treat diseases. Conversely, unsuitable sleeping posture would do harm to human health. Therefore, it is of great significance to accurately and quickly distinguish human sleeping posture for improving sleep quality and formulating disease treatment decision. Therefore, in order to further enhance the accuracy of sleeping posture recognition, we used a Infrared night vision camera to collect sleeping data, and proposed a deep learning-based predictive model (DL-model). The DL-model introduced the attention mechanism to and capture the characteristics of significantly related to the target task and reduce the expression of redundant characteristics. Meanwhile, considering that when the blanket is covered, some human features would be blocked, we took whether the blanket covered as auxiliary task. The experimental results showed that the average accuracy of the DL-model reaches 94.62%, which was better than the other state-of-art deep learning methods. Compared with signal task learning, the DL-model using multi-task learning performed well. Also, the DL-model we proposed in this study also has a good performance in the auxiliary task–whether to cover the blanket. To some extend, it confirmed the feasibility of using the monitoring camera for sleeping posture prediction, which is of great significance for realizing the real-time prediction and sleep quality assessment at night.