Polarization Weighting Based Improved HRNet for Preterm Infants Pose Estimation

Accurate pose estimation of preterm infants in neonatal intensive care units (NICUs) plays an important role in early screening of diseases. The monitoring of subtle limb movements caused by pain, abnormal muscle tone, neurological damage, etc., can provide valuable diagnosis clues for healthcare professionals. However, poor generalization performance is achieved when applying pose estimation algorithms of adults directly due to the significant differences of body proportions between adults and preterm infants, as well as the complex backgrounds of NICUs. To address this issue, an improved HRNet network is proposed by incorporating a polarized self-attention (PSA) to mitigate interferences, a novel adaptive fusion module and a multi-branch channel output module for extracting inter-joint correlations. Experiments on real-world preterm infant video data demonstrate the effectiveness of the proposed method.