Real-Time Pose Estimation of Preterm Infants Using Depth Images

Early diagnosis of neurodevelopmental disorders in infants relies on accurate analysis of spontaneous movements. Achieving this requires fast and precise pose estimation methods tailored to infant-specific anatomy and motion. This study evaluates several pretrained YOLOv11-pose models for pose estimation in depth video recordings of preterm neonates and infants using the open source babyPose data set database. The fastest model (YOLOv11n-pose) has a inference time of 0.007 seconds. Considering a previously proposed data split without subject-wise separation between training and testing data, the most accurate model (YOLOv11m-pose) has a median root mean squared distance (RMSD) of 2.15. The median Dice Similarity Coefficient (DSC) and Recall (R) of the joints are 0.85 and 0.86, while the median DSC and R of the joint connections are 0.90 and 0.91. Considering a subject-wise separation of training and testing data, the results noticeably degrade, e.g. to a median DSC and R of the joints of 0.79 and 0.81, while the median DSC and R of the joint connections are 0.75 and 0.79. The present work demonstrates a fast and, copared to the literature, accurate approach to depth-based pose estimation in preterm neonates and infants paving the way for automated movement analysis as a clinical tool for early detection of developmental impairments. Particularly in semiautomated settings where subject-specific annotations can be provided, the results are convining. Regarding the abilities to generalize, more work is required to improve the results.