Deep transfer learning for video-based detection of newborn presence in incubator.

Preterm newborns are prone to late-onset sepsis, leading to a high risk of mortality. Video-based analysis of motion is a promising non-invasive approach because the behavior of the newborn is related to his physiological state. But it is needed to analyze only images where the newborn is solely present in incubator. In this context, we propose a method for video-based detection of newborn presence. We use deep transfer learning: bottleneck features are extracted from a pre-trained deep neural network and then a classifier is trained with these features on our database. Moreover, we propose a strategy that allows to take advantage of temporal consistency. On a database of 11 newborns with 56 days of video recordings, the results show a balanced accuracy of 80%.