Video-Based Face and Facial Landmark Tracking for Neonatal Vital Sign Monitoring

Abstract

This paper explores automated face and facial landmark tracking of neonates, for the purposes of vital sign estimation. Utilising a publicly available dataset of neonates in the clinical environment, 25 videos were annotated. Face and facial landmarks (i.e. eyes and nose) tracking are then assessed. Additionally, the identification and tracking of the neonate's forehead and cheeks are purposed, as they are ideal regions of interest for vital sign estimation. Tracking of the face produced an average overlap score of 93.0%. Tracking of the eye and nose landmarks produced mean normalised errors of 0.026 and 0.019 respectively. The cheek region of interest could be effectively identified and tracked, whereas the forehead region of interest identification was incorrect 16% of the time.