Development of Domain Based Model for Sleep Posture Prediction at Night

Sleeping posture has a certain correlation with people's sleep health. Accurate assessment of sleeping posture is of great significance for sleep quality monitoring. As is known to all, the sleeping environment in most cases is low light or even dark. Surveillance cameras with infrared night vision function, as a non-invasive monitoring tool, could be used to identify sleeping postures during the day and at night. However, most current research was based on analysis and modeling of sleeping posture images under good lighting conditions. there was little studies proposed a nighttime sleep posture data set using surveillance cameras. Therefore, in order to make the established model have better performance in sleeping posture recognition, in this study, we collected sleeping posture data both in daytime and nighttime scenes. In addition, we proposed a sleeping posture recognition model based on domain adaptation and attention mechanisms, combining on existing public datasets and our dataset. We used a domain adaptation algorithm to transfer daytime sleeping posture features to nighttime sleeping posture images, and we also introduced an attention mechanism to better capture the effective features in the sleep posture images and enhance the generalization and stability of the model. Experimental results showed that our model proposed in this study could achieve accurate nighttime sleeping postures when there is a lack of nighttime sleeping posture data.