3D Point Cloud-Based Visual Prediction of ICU Mobility Care Activities

Intensive Care Units (ICUs) are some of the highest intensity areas of patient care activities in hospitals, yet documentation and understanding of the occurrence of these activities remain sub-optimal due in part to already-demanding patient care workloads of nursing staff. Recently, computer vision based methods operating over color and depth data collected from passive mounted sensors have been developed for automated activity recognition, but have been limited to coarse or simple activities due to the complex environments in ICUs, where fast-changing activities and severe occlusion occur. In this work, we introduce an approach for tackling more challenging activities in ICUs by combining depth data from multiple sensors to form a single 3D point cloud representation, and using a neural network-based model to reason over this 3D representation. We demonstrate the effectiveness of this approach using a dataset of mobility-related patient care activities collected in a clinician-guided simulation setting.